Cutting Payload aka Packet Trimming—the CRISPR-Cas9 of Network Flow Control?

(Edited to add more pointers to implementations at the end— SL.)

One of my tasks in the (backbone) network team at Switch involves identifying and following "trends" outside our bubble that might somehow end up affecting our business. A current trend with possible impact is "AI"—maybe you have heard of it—which today means mostly machine learning using (deep) artificial neural networks, and in particular large language models (LLMs).

There are already noticeable changes in how datacenters are built to accommodate LLMs—check out what OCP has been doing in the past few years. These changes include (datacenter) networks, which must now support the needs of large GPU clusters, in particular for LLM training. To use those high-performance—and high-cost—GPUs effectively, one wants to be able to copy large amounts of data between GPUs anywhere in the cluster, which today might contain 32768 GPUs or so, with some striving up to a scale of millions. These transfers typically use remote memory-to-memory transfer without CPU involvement (RDMA). There are significant efforts to make this work over some sort of standardized Ethernet-based network. This could be done with ROCE (RDMA over Converged Ethernet) (v2), where the "Converged" hints at some modifications beyond traditional Ethernet.

Some techniques were already considered for addition to Ethernet for datacenter networks before the LLM hype wave, for example PFC (priority flow control) to support "lossless" operation of the network for specific subsets of traffic, or novel ECMP (equal-cost multipath) approaches that don't necessarily preserve per-flow ordering (using "spraying" instead).

Recently I was excited when I heard about another technique to be added: packet "trimming". It reminded me of a great (video) presentation by Mark Handley (UCL) from SIGCOMM 2017. I'll add the link later. For now I'll refer to this work as "the NDP paper".

I remember being impressed when I saw the video, and I also found that it was well received. Just having a paper accepted at SIGCOMM is one of the highest achievements for (Internet-affine) network researchers, and this paper was awarded best paper that year. I also remember thinking something along the lines of... "wow, that was an awesome presentation and must have been the talk of the conference... but what are the chances of this being implemented in production networks within my lifetime?". Focused on "classical" Internet backbones and enterprise networks, I thought that this was just a bit too revolutionary/"disruptive" to have much of a chance in our environment, which has shown strong resistance to seemingly much simpler changes (AQM, ECN, making addresses wider...). But 7–8 years later and thanks to the investment craze kicked off by ChatGPT, maybe the time is ripe!

One thing I was wondering about was whether the NDP paper was the first to introduce "trimming". Upon re-watching the video after several years (and understanding a bit more than at previous attempts :-) I caught Mark referring to previous work from Tsinghua as the inspiration for the trimming. A bit of searching turned up this NSDI 2014 presentation of the paper "Catch the Whole Lot in an Action: Rapid Precise Packet Loss Notification in Data Centers" by Peng Cheng (then a Ph.D. student at Tsinghua University, now at Microsoft Research Asia) et al. It includes a nice introduction where Peng Cheng talks about cooking and his use of the scissors as a multipurpose tool, suggesting that his "cut payload" approach could be similarly broadly useful. (A bit like the CRISPR "gene scissors" in bioengineering today, no?) Maybe Peng Cheng et al.'s work could be considered for a "test of time" award.

So where are we now, in 2025? Looks like concepts from the NSDI 2014 and SIGCOMM 2017 papers (trimming, spraying...) will be added to the Ultra Ethernet standard. It is not clear to me how the trimming capability will be used for congestion management. Maybe not at all in the initial standard?

Thinking a bit further: Assuming that trimming becomes a standard function in data center switches, could we use it in other network contexts? For example, to reduce "bufferbloat" for broadband users, or generally in the wide-area Internet... If people see potential for this, then I guess some work (including standardization work, presumably in the IETF) would be needed to ensure that trimmed packets could safely be sent over the Internet, and on Internet-deployable transport protocols (based on something like NDP?) that can make use of them.

My intuition says that there is some potential there, and if done well, this novel idea could become a useful addition to the Internet.

But one assumption that justifies trimming is that "serialization delay" (i.e. the time it takes to put the bits of a packet on the wire) is significant relative to "propagation time" (the time it takes the packet to travel through the network). This can be true for networks that are relatively small in geographical size—for example within a data center—or for networks that have relatively "slow" (low-bitrate) links; or both, of course! (There is also a dependence on packet size, but for now let's assume that the maximum packet size is similar for all networks, between ~1500 and ~9000 bytes. This is the case in practice today, although it would be interesting to study in how far trimming could make large-MTU networks more viable.)

So assuming that trimming makes sense in a DC network where end-to-end round trips take on the order of a microsecond (100m distance) and link speeds are on the order of 100Gb/s, it should make just as much sense on a network with one millisecond RTT and links of 100Mb/s, or one with 100ms RTT and links of 1Mb/s. OK, such delay/rate combinations aren't frequently seen in classical "research & education networking" anymore these days. But some interesting link technologies used at the network "edge" may sometimes exhibit such low rates—Wi-Fi when the station is far away from the base, or low-energy short/medium-distance wireless.

My gut feeling is that packet trimming may become commonplace in data center networks, provided that the AI "summer" continues for a few years. It may disappear again because the complexities (integrating it with higher-level protocols) may outweigh the gains. But it's encouraging that such "outside-the-box" ideas can gain traction in the market—when the stars align just right.

Oh yeah, here is the promised link to Mark Handley's presentation: Re-architecting datacenter networks and stacks for low latency and high performance, SIGCOMM 2017 proceedings. See the "supplementary material" section for the video.

Known Packet Trimming Implementations

Cumulus Linux ≥5.14 supports Packet Trimming on NVIDIA Spectrum chipsets (Release Notes).

Arista EOS ≥4.35 supports Packet Trimming on (Tomahawk 5-based?) 800G switches (Release Notes).

HPE Juniper supports Packet Trimming (called DCN, Drop Congestion Notification) on Tomahawk 5-based switches (Configuration Guide, LinkedIn post).

Broadcom's MRC (Multi-Path Reliable Connections) on Tomahawk 5/6 and Thor Ultra 800GE NICs uses source-controlled "spraying" and packet trimming (blog post by N. Vaidya, C. Raicu, and E. Spada), May 2026.

Google Code Jam 2023 in Lisp

Since 2009, I have occasionally participated in Google's Code Jam competitions, usually programming in Lisp. Google decided to discontinue it along with their other coding competitions. So this year there was a "Farewell Round" held on a single day (yesterday), actually four rounds (A, B, C, D) held during the same long (four-hour) timeslot. Because I didn't feel very competitive or confident, I mainly tackled the problems from the "A" round, assuming that those were the easiest ones.

Thanks to the generous timeslot, I managed to solve all five problems in Round A.  You can find them on the gcj2023 project on my personal GitLab server.

Colliding Encoding: The task is to check whether there are collisions (duplicates) in a list of word after encoding with a lossy cipher function (mapping 26 alphabetic letters to 10 decimal digits). I map the strings, sort them, and walk through the result comparing neighboring entries—iff there's a match, then there are collisions. Code: a/1.lisp

Illumination Optimization: Given a set of lampposts at positions X_i along a road, how many lightbulbs of illumination radius R are needed to light the entire stretch of road from 0 to M? You go through the lamppost positions in order and skip those that are already illuminated by the previous lamp. My solution (a/2.lisp) tries both directions, in case one is "more optimal" than the other :-)

Rainbow Sort: On first read, I found it a bit confusing, because it talks about colors, card positions, and ordering numbers, all represented as integers; so I put it off until after the others. In the end it turned out to be quite easy. My solution (a/3.lisp) is a simple recursive function using a hash table to store colors that have already been seen.

ASCII Art: I quickly figured out that there must be a solution that involves just computation and no searching or construction. But it took me quite a bit of time (easily more than an hour) and trial/error until I removed all off-by-one errors. The solution (a/4.lisp) ended up simple enough, and solved all sets on the first attempt.

Untie: Here again, I quickly figured out that I can simply ignore all sub-sequences that have no repetitions. For every repeated sequences of length N, you can always fix the tie by flipping ⌊N/2⌋ symbols. I convinced myself that this is always possible without creating additional collision, thanks to the fact that you have two symbols to choose from. The trickiest part is that we're dealing with a circle rather than a sequence with start and end. My approach was to start at a point where the symbol changes—if there is no such point, then this is a special situation of a circular sequence of the total length of the sequence, and in this case actually you need to flip ⌈N/2⌉ symbols. Other than that, my recursive solution (a/5.lisp) looks relatively simple, and solved all cases on the first attempt.

By solving all problem sets in time, I ended up in rank 734 for that round, which made me quite happy! With some (but too little) time remaining, I tried my hand at two problems from more difficult rounds, but didn't manage to solve either of them in time.

Game Sort (Round C): This didn't look that hard, but my code consistently solved only the tiny sample set, and failed at even the small set on all my attempts. Maybe I take my broken code (c/1.lisp) up and try to find and fix what's wrong with it—or probably rather with my initial understanding of the problem and its possible solution.

Spacious Sets (Round B): My first iteration of code solved the small set, but ran into a timeout on the large set. I had to revise it to avoid quadratic complexity. I first thought of "memoization" to optimize away repeated iterations; in the process of implementing that, I discovered a nice linear initial pass (which has to be done in both directions—the problem has some structural similarities with Illumination Optimization above) that allowed the computation for a given pivot element to be done in constant time (adding two pre-computed numbers and 1). To my surprise, even with this probably very close to optimal solution, my program ran out of time! It turned out that I had another O(N²) issue in an innocuous part of the code that simply constructs a mapping from indexes in the original unordered sequence to the position in the sorted sequence that I use for the computation. I had to replace the trivial code using POSITION with my own FAST-POSITION trivially implementing binary search on the sorted output vector. The code (b/3.lisp) doesn't look great, but it works! I only really started after the end of the competition, though; otherwise I might have gotten four points for the small set...

So that was Google Code Jam's Farewell Round! Thanks a lot to Google for organizing those events over twenty years, and especially to the many Googlers for coming up with great problems, ironing them out, and providing tons of challenging data sets! I had a lot of fun—along with quite a bit of frustration; but that goes along with the challenge, I guess.

AI/ML Hype: If even Google doesn't get it right, then...?

 I'm a happy and frequent user of Google Photos, which I use to back up, share, and edit the many photos that I take on my smartphone.

This morning, the mobile app greeted me with a new feature: It suggested that six of my photos were incorrectly oriented, and offered me to fix that by rotating them.

I checked, and out of those six photos, five were totally fine as they were. (I'm a bit pedantic and usually fix any wrong orientation quickly by hand, which Google Photos makes quite easy.)

The sixth was indeed misoriented, but the rotation suggested by the tool was also wrong.

First I laughed. If Google—of all people!—sends me twelve predictions via a highly popular app, and gets eleven of those wrong, how can anybody really expect that Artificial Intelligence/Machine Learning will solve real problems and eventually pay back all the investments being made in it? And note that the problem class here is basically image classification, which is one of the few narrow domains where ML has been particularly effective.

And then I dutifully corrected Google's mispredictions by rotating the proposed images until they were correct (correct again, in five of the six cases). So maybe if a few million other nice/gullible people help poor Google out, then maybe one day, it will actually become helpful even to pedants like me.

This suddenly recalled a question Prof. Mireille Hildebrandt asked at a large EU event last week ("Leading the Digital Decade", video):

What is a digitally skilled population? Is it a population capable of using AI and other digital systems? That's the usual way to talk about citizens: users. Or is it, perhaps, a population that is open to be used by these systems as data engines?

Google Code Jam 2020 in Lisp—Round 1C

Well, that was slightly embarrassing and frustrating...

After a relatively promising attempt at round 1B, I entered round 1C yesterday as the last chance to advance to round 2. Again, three nice problems, with the third one looking too hard for me at first sight, at least for the general solution. My "plan" after reading the problem descriptions was roughly: Solve problems 1 and 2 and, time permitting, submit a solution for the easy first input set of problem 3. That would have been a good plan, except I took a bit too long to code problem 1, and I got completely stuck trying to debug my Lisp solution for problem 2.

Problem 1: Overexcited Fan

The problem turned out to be even easier than I first thought. I should have spent more time designing the algorithm before starting to code it. I had an OK solution, but it was more complex than necessary and took me 53 minutes. That's much longer than it should have, given the overall time limit of 2 hours 30 for the three tasks. After the round, I simplified this to a more elegant solution.

Problem 2: Overrandomized

Here I had the correct inspiration to work with letter frequencies, in particular first-digit letter frequencies. Unfortunately my Lisp solution would always result in "RE" (runtime error) when submitted, even though it ran fine both against the sample data set and against some additional data sets that I have generated. At this point I ran out of time.

After the round, I tried for some time to get my Lisp program working, but without any possibility to get diagnostic output, I didn't find where the problem lied. I even started from scratch using a different implementation technique, but the result was the same: My program would run fine against the sample and my self-generated tests, but throw "RE" when faced with the competition/training data. Very frustrating. If anyone finds the error(s), please let me know!

At one point I reimplemented my solution in C++, and it successfully solved all data sets as soon as I got it to compile. I then "backported" it to C, which simplified it further (in the C++ version I used both "traditional" arrays and "library" vectors—for sorting—which makes that version quite ugly.

Problem 3: Oversized Pancake Choppers

After reading the problem description, I assumed (probably correctly) that a full solution would beyond my capabilities for solving, at least under this time pressure. But the easy set seemed quite tractable: If there are only 2 or 3 diners, then the set of solution possibilities is quite bounded. Indeed it took me only (competitive coders can laugh now) about 15 minutes to write a stupid solution that worked for this limited case. But that's academic because I had run out of time in problem 2, so this happened after the round was over.

Conclusion

With only the first and easiest problem solved in time, I ranked in the 8000s, much below my rank in 1B—and hopelessly below the cut-off at 1500.

In problem 1 and (my constrained/stupid solution for) problem 3 I had no problems with Common Lisp as a coding language. But I failed to get my problem 2 Lisp solution working, while later attempts in C++ and C worked right away (and showed that I had analyzed the problem correctly).

Google Code Jam 2020 in Lisp—Round 1B

Although I had quite a bit of trouble surviving this year's qualification round, I was allowed to participate in online round 1, which, as usual, is held as three different sub-rounds to accommodate participants from various time zones.  The first (1A) was in the middle of the night in my time zone, so I skipped it and didn't even get around to reading the problems.  But today I attempted  round 1B.  The problems were again very cute and interesting, and while I struggled with the limited time, I ended up being quite satisfied, even though I missed qualification this time.

Problem 1: Expogo

I wrote a relatively simple implementation using search with memoization (caching) and a bit of pruning of implausible search regions, and that was sufficient for all three data sets.  According to the after-round summary, there's also a constructive solution.  I had suspected so, but couldn't find it.

Problem 2: Blindfolded Bullseye

This is a nice search problem that involves some geometry.  I think I started with a reasonable search strategy, but then ran out of time when it came to doing the geometry... So I gave up trying to find a "good" solution, and coded the trivial brute-force solution for solving the easiest data set.  That was worth only 3 points, but those 3 points improved my final ranking by more than 1'000.

Problem 3: Join the Ranks

This didn't look too impossible, but given the high amounts of points awarded for this, I suspected that this would be beyond my skills and decided to focus on the other two problems.

Conclusion

With a full solution for problem 1, a half(?)-done good solution for problem 2, and the few extra points for the trivial data set of problem 2, I finally ranked #1996 [edited for final result].  As only the first 1500 contestants qualify for round 2, this was not sufficient.  But I'm quite happy with how it went, and am looking forward to round 1C, about two weeks from now.  And unlike during the qualification round, I didn't run into any issues with coding in Lisp—even though I haven't been using it professionally for several years now.

Google Code Jam 2020 in Lisp—Qualification Round

This year I registered for Google Code Jam again. The new platform supports Common Lisp (SBCL) again, so I started writing my solutions for the Qualification Round in SBCL. Although I managed to qualify for Online Round 1, it didn't work as well as I'd hoped.
The Qualification Round this year had five problems, which took me a while to notice—I had to scroll horizontally to see the fifth problem. I attempted all five, but only managed to write working solutions for three of them.

Problem 1: Vestigium

That was an easy one, but for some reason I didn't manage to code it correctly in Common Lisp. Even though my code looked totally straightforward, it only produced an RE (Runtime Error) when run against the data set. Finally I rewrote the solution in C++, and that worked on first try.

Problem 2: Nesting Depth

The task was to generate appropriate numbers of correctly nested parentheses. Obviously this type of problem is familiar to Lispers, and this time my straightforward code actually worked. My sloppiness resulted in an invalid attempt, as I initially left out the (solve) call at the end of the script.

Problem 3: Parenting Partnering Returns

It occurred to me relatively early that the problem maps to the problem of partitioning (two-coloring) the interval graph. However, all my attempts of coding this simple graph traversal algorithm in Common Lisp ended in REs (Runtime Exceptions) again. Very frustrating!

Problem 4: ESAb ATAd

This was my favorite problem, and I came up with a simple and near-optimal solution pretty quickly. It would have worked the first time, had I not (again) forgotten to add the correct (solve) call after the development phase.

Problem 5: Indicium

That one was definitely too hard for me! After the deadline, I read the analysis and took it up again. I managed to code a super-efficient method to generate latin squares with arbitrary possible traces at will, based on "circulating" matrices. The only cases that I cannot solve in this way are odd-sized with traces of is n+2 or n²-2.

Conclusion

Finally I advanced with 49 points (from 100, with 30 required), and ranked #6318 of about 44000 participants. My Lisp development and graph algorithm skills have become a bit rusty, and I'm not very optimistic about being able to survive Online Round 1.

Compiling Emacs on Mac OS against X11 libraries

Motivation

My place of work has been using Apple as the main, and now basically the only, supported platform for office use. As a long-time (GNU) Emacs user and occasional contributor, I usually compile GNU Emacs from the development (master) sources, which are now thankfully available over Git.
I already feel bad for using a proprietary OS rather than something based on GNU/Linux or another free or at least open source system. For this and other reasons, I try to at least use system libraries from the free/open source community rather than Apple's, for example for the GUI.  So rather than Apple's Carbon UI, I build Emacs against X11 libraries. Previously I could do this with the Gtk toolkit, but the maintainers of the Mac OS port of Gtk I was using stopped supporting its X11 drivers. So I reverted to the older "Lucid" toolkit with X11. (Yes, maybe I'm weird to cling to the X Window System in spite of the obstacles... that's probably because X was an important part of my socialization to Free Software and distributed systems.)

Howto

The prerequisites are

• a bunch of packages from Homebrew (not all of which may be relevant):
• autoconf autogen automake cairo d-bus dbus fontconfig freetype gcc gdk-pixbuf gettext git glib gmp gnutls gobject-introspection graphite2 harfbuzz imagemagick intltool ispell jansson jpeg json-glib libffi libpng librsvg libtasn1 libtiff libtool libunistring libxml2 little-cms2 netpbm nettle openjpeg openshift-cli osinfo-db osinfo-db-tools pcre pcre2 pkg-config readline shared-mime-info texinfo zlib
• an X11 (XQuartz) installation with development headers
• The Xcode development environment with its CLI tools installed.

I clone Emacs into /var/tmp/emacs/emacs according to the instructions on the Emacs from Git page of the Emacs Wiki. Then I mkdir /var/tmp/emacs/gbuild, cd there, and run

../emacs/configure --verbose \
  --with-x --with-x-toolkit=lucid --with-ns=no \
  --without-makeinfo \
  LIBXML2_CFLAGS=-I/usr/local/opt/libxml2/include/libxml2 \
  LIBXML2_LIBS='-L/usr/local/opt/libxml2/lib -lxml2' \
  --with-jpeg=no --with-gif=no --with-tiff=no \
  --x-libraries=/usr/local/opt/freetype/lib:/usr/X11/lib \
  --x-includes=/usr/local/opt/freetype/include:/usr/X11/include \
  --with-xpm=no \
  PKG_CONFIG_PATH=/usr/local/lib/pkgconfig:/usr/X11/lib/pkgconfig
This gives me the following configuration:
Configured for 'x86_64-apple-darwin18.6.0'.

  Where should the build process find the source code?    ../emacs
  What compiler should emacs be built with?               gcc -g3 -O2
  Should Emacs use the GNU version of malloc?             no
    (The GNU allocators don't work with this system configuration.)
  Should Emacs use a relocating allocator for buffers?    no
  Should Emacs use mmap(2) for buffer allocation?         no
  What window system should Emacs use?                    x11
  What toolkit should Emacs use?                          LUCID
  Where do we find X Windows header files?                /usr/local/opt/freetype/include:/usr/X11/include
  Where do we find X Windows libraries?                   /usr/local/opt/freetype/lib:/usr/X11/lib
  Does Emacs use -lXaw3d?                                 yes
  Does Emacs use -lXpm?                                   no
  Does Emacs use -ljpeg?                                  no
  Does Emacs use -ltiff?                                  no
  Does Emacs use a gif library?                           no
  Does Emacs use a png library?                           yes -L/usr/local/Cellar/libpng/1.6.37/lib -lpng16 -lz
  Does Emacs use -lrsvg-2?                                no
  Does Emacs use cairo?                                   no
  Does Emacs use -llcms2?                                 yes
  Does Emacs use imagemagick?                             no
  Does Emacs support sound?                               no
  Does Emacs use -lgpm?                                   no
  Does Emacs use -ldbus?                                  yes
  Does Emacs use -lgconf?                                 no
  Does Emacs use GSettings?                               no
  Does Emacs use a file notification library?             yes (kqueue)
  Does Emacs use access control lists?                    yes
  Does Emacs use -lselinux?                               no
  Does Emacs use -lgnutls?                                yes
  Does Emacs use -lxml2?                                  yes
  Does Emacs use -lfreetype?                              yes
  Does Emacs use HarfBuzz?                                yes
  Does Emacs use -lm17n-flt?                              no
  Does Emacs use -lotf?                                   no
  Does Emacs use -lxft?                                   yes
  Does Emacs use -lsystemd?                               no
  Does Emacs use -ljansson?                               yes
  Does Emacs use -lgmp?                                   yes
  Does Emacs directly use zlib?                           yes
  Does Emacs have dynamic modules support?                no
  Does Emacs use toolkit scroll bars?                     yes
  Does Emacs support Xwidgets (requires gtk3)?            no
  Does Emacs have threading support in lisp?              yes
  Does Emacs support the portable dumper?                 yes
  Does Emacs support legacy unexec dumping?               no
  Which dumping strategy does Emacs use?                  pdumper
Then I can build Emacs using
make bootstrap
This produces an intermediate "temacs" binary referencing many shared objects:
: 1leinen@macsl[leinen]; objdump -macho -dylibs-used /var/tmp/emacs/gbuild/src/temacs
/var/tmp/emacs/gbuild/src/temacs:
 /opt/X11/lib/libpng16.16.dylib (compatibility version 43.0.0, current version 43.0.0) /usr/lib/libz.1.dylib (compatibility version 1.0.0, current version 1.2.11) /opt/X11/lib/libXaw3d.8.dylib (compatibility version 9.0.0, current version 9.0.0) /opt/X11/lib/libXmu.6.dylib (compatibility version 9.0.0, current version 9.0.0) /opt/X11/lib/libXt.6.dylib (compatibility version 7.0.0, current version 7.0.0) /opt/X11/lib/libSM.6.dylib (compatibility version 7.0.0, current version 7.1.0) /opt/X11/lib/libICE.6.dylib (compatibility version 10.0.0, current version 10.0.0) /opt/X11/lib/libXext.6.dylib (compatibility version 11.0.0, current version 11.0.0) /opt/X11/lib/libX11.6.dylib (compatibility version 10.0.0, current version 10.0.0) /opt/X11/lib/libX11-xcb.1.dylib (compatibility version 2.0.0, current version 2.0.0) /opt/X11/lib/libxcb.1.dylib (compatibility version 3.0.0, current version 3.0.0) /opt/X11/lib/libXft.2.dylib (compatibility version 6.0.0, current version 6.2.0) /opt/X11/lib/libXrender.1.dylib (compatibility version 5.0.0, current version 5.0.0) /usr/local/opt/dbus/lib/libdbus-1.3.dylib (compatibility version 23.0.0, current version 23.11.0) /opt/X11/lib/libXrandr.2.dylib (compatibility version 5.0.0, current version 5.0.0) /opt/X11/lib/libXinerama.1.dylib (compatibility version 2.0.0, current version 2.0.0) /opt/X11/lib/libXfixes.3.dylib (compatibility version 5.0.0, current version 5.0.0) /usr/local/opt/libxml2/lib/libxml2.2.dylib (compatibility version 12.0.0, current version 12.9.0) /usr/lib/libncurses.5.4.dylib (compatibility version 5.4.0, current version 5.4.0) /usr/local/opt/freetype/lib/libfreetype.6.dylib (compatibility version 24.0.0, current version 24.1.0) /opt/X11/lib/libfontconfig.1.dylib (compatibility version 11.0.0, current version 11.2.0) /usr/local/opt/harfbuzz/lib/libharfbuzz.0.dylib (compatibility version 20504.0.0, current version 20504.0.0) /usr/local/opt/gnutls/lib/libgnutls.30.dylib (compatibility version 55.0.0, current version 55.0.0) /usr/local/opt/little-cms2/lib/liblcms2.2.dylib (compatibility version 3.0.0, current version 3.8.0) /usr/local/opt/jansson/lib/libjansson.4.dylib (compatibility version 16.0.0, current version 16.1.0) /usr/local/opt/gmp/lib/libgmp.10.dylib (compatibility version 14.0.0, current version 14.2.0) /usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 1252.250.1)
...and an src/emacs that hopefully doesn't crash. Oops... why am I saying that?
Well, yesterday I used a simpler method that didn't specify /usr/local/opt/freetype explicitly anywhere, but relied on autoconf/pkgconfig to find it. Unfortunately this created a "DLL hell" situation where an older version of the freetype dynamic library from /opt/X11/lib was used instead of the newer on in /usr/local/opt/freetype/lib, leading to crashes when Emacs tried to use certain fonts. Thanks to YAMAMOTO Mitsuharu for helping me find the error in my old build process!

Google Code Jam 2019 in Lisp (again) - Qualification Round

Like most years, I registered for Google Code Jam again, because it's fun. The qualification round is usually easy for me to survive, but presents an opportunity to de-rust the brain, get used to coding and interacting with the competition platform.
Google introduced a new platform last year with a big change: Previously you got input and submitted output for checking, now you submit code, and Google runs it against input to check whether it generates the correct results. I suspect that a major driver for this change was the desire to curb cheating. The new platform also allows "interactive" problems, where your code interacts (via stdout/stdin) with a "judge" process, which is cool.
One sad effect of last year's move to the new platform was that I could no longer use Lisp—only a couple programming languages were supported. Thus I mostly used C last year.
This year I thought it was the same, and coded the first two problems, Foregone Solution and You Can Go Your Own Way, in C. As I started the third, I somehow noticed that more programming languages are supported this year, including Lisp (SBCL) and Clojure. So I changed to Common Lisp for problems three and four, Cryptopangrams and Dat Bae.
After fighting a little with the way the platform runs Lisp programs (I hadn't even known about SBCL's --script option), I managed to submit working versions for all problems.
When the ratings came out, it turned out that I had gotten the two difficult problems right including the (hidden) large datasets, but for the two easy problems, my results for the large inputs were bad: Wrong Answer for Foregone Solution and Runtime Error for You Can Go Your Own Way.
For the first, I had used strncpy() with a limit of MAXDIGITS, and forgot that if there are actually MAXDIGITS characters to copy, then strncpy() won't copy the terminating NUL, so I had to write that NUL byte to the end of the string, which I had failed to do.
For the second, I didn't read the problem description well, and thought that the problem size was limited to a 1000x1000 maze, where in fact it was 50000x50000 for the large set. Because I statically assigned vectors based on the maze edge length, this caused overruns. Easy to fix by changing a #define from 1000 to 50000.
Both bugs would have been unlikely in Lisp, where strings don't have to be NUL-terminated and dynamic allocation of data structures is more natural than fixed allocation.
Anyway, I lost 1+10 points to these stupid bugs, which gave me an overall score of 89. Of the more than 35500 participants, only one other person and I had this particular score. My rank was 1147, which is somewhat encouraging. But I was lucky to find good and easy-to-program approaches for each of the problems, and that will become (much) harder in the coming rounds.
For Foregone Solution, I noticed quickly that xx4x44x can be trivially split into xx3x33x and 0010110. For You Can Go Your Own Way, it took me a bit longer to come up with the "complement" solution, i.e. when Lydia has walked SSESEESE, then I can simply walk EESESSES and be sure that we both arrive at the same destination (because the maze is square) and never make the same move. With Cryptopangrams, I quickly had the correct intuition that I should compute the GCD of adjacent cryptosymbols to extract the common prime, and hacked together a version that would solve the two simple sample cases. But as I repeatedly failed the smaller input sets, I noticed that there are complications when symbols are repeated. This was when I resorted to pencil and paper—it would probably have been a good idea to start this earlier!—, and it took me a while to turn this into code. But when I thought I had it right, the system still wouldn't accept my solutions! That's when I found out that I have to actually call the (solve) function at the end of my program—doh!
For Dat Bae, I implemented the interaction with the Python-written sample judge using a subprocess, which allowed me to run local tests. The easier set (F=10 guesses for broken bits in up to 1024 total bits) was solved by numbering all 1024 bits. It took me a break to find that I could easily bring the number of guesses down to F=5 due to the constraint that not more than 15 bits are broken: This allowed me to use a smaller pattern (e.g. 0...31) and repeat it, without introducing any ambiguities.

You say "the internet", I say "the Internet"...

My friend Nick Feamster recently twote something that resonated with me:

I still bristle whenever I read Internet with a lowercase "i".  The Internet is one specific instance of an "internetwork". People who write editorial style guides for tech should be required to take a networking class, or at least a history of networking class.

(The perspicacious reader will note that this happened after the 280calypse)
Having been involved for many years with the Internet, and the Internet Engineering Task Force, and the Internet Society, my first thought was "YES! Exactly this happens to me all the time."
But then I read the whole tweet, and much of the discussion, and found that I agree even more with Heather E. Merrick, who responded:

The AP Style Guide used to call for this but they grew up in 2016.

...especially after she countered Nick's remark that AP's editors "should clue up on history, and the use of proper nouns" with this explanation:

They used to, and then they wised up to the reality that language is not in the hands of any one person or industry but The People, who enjoy using a lowercase "i," hence the update to the Stylebook.

 Ratul Mahajan put it more succinctly (adding a friendly health tip):

inevitable success disaster

So you built the "Internet", and other people write about the "internet". My pro tip: Take it as a huge compliment: The thing you built has become an important part of their environment. Too vital to be considered some brand name anymore. But of course if you want to be sad about people "minimizing" your invention, that's your choice. Here's a kleenex for you!
Yes, I'll continue to write Internet with a capital I. Just like I write the name of my employer as "SWITCH" as we're told to, even though everybody always does it wrong and writes "Switch". But I stopped getting mad at those people, it's just not helpful. Instead I thank them for letting me continue to use the "correct" spelling.
When they tell me I have to stop, I'll just turn their internet off, that'll show them.

"...let's turn the turning off off", OK?

Google Code Jam 2017 in Lisp: Round 1B/1C

Online round 1 was done on three dates again this year. The first (1A) fell in the middle of the night so I skipped it.

Round 1B Problem A: Steed 2: Cruise Control

I attempted the second, 1B, but I was distracted by important household chores. That's also why I didn't make any sketches. The first problem was easy enough, and I found a solution while fixing a meal for a kid who'd just returned from boy scout camp. The basic idea is that the horses all queue up behind the slowest one, so the arrival time is the latest arrival time at the destination point for all horses. The solution (A.lisp) worked fine on the first attempt of the small dataset, and also on the large one.

Round 1B Problem B: Stable Neigh-bors

The large dataset, where the unicorn's mane can have mixed colors, looked significantly more complex than the small one, where there are only three colors. I attempted a straightforward greedy algorithm, but somehow got it wrong. Because I didn't feel in a good state of mind, I decided to give up at that point. Otherwise I would have attempted the small dataset of Problem C: Pony Express. The "large dataset" cases of both problem B and problem C looked just too difficult to me.

Round 1C Problem A: Ample Syrup

A quick analysis suggested to me that the radius of the pancakes isn't relevant, because only the bottom (largest) one "counts", so I need to pick the pancakes with the largest lateral surface. Of course that reasoning was wrong, because we also need to pick which will be the bottom pancake. This error cost me two bad attempts, but I managed to reuse most of the initial code by adding a loop around it that tries various base pancakes. The code is in A.lisp and, again, worked fine for both the small and the large dataset.

Round 1C Problem B: Parenting Partnering

The optimization problem looked a bit tricky (but doable) in the general case, but the "small dataset" case was very specialized and easy enough to attack. There are only 1 or 2 activities in total. Either each partner has one, or one partner has all 1 or 2. If each partner has one, then the optimal solution always involves two exchanges. If one partner has the only one activity, the solution is even more trivially two exchanges. If one partner has (all) two activities, then the optimal solution is two if those activities can be arranged on the same "half-circle" (12 hours or 720 minutes) of the day, otherwise four. I coded these cases in B.lisp and was successful on the small dataset. In hindsight I'm surprised that it took me so long, more than 45 minutes between my submissions for A and for B-small.

Round 1C Problem C: Core Training

Next, I decided to try the small dataset of problem C. While I thought B-large might be in reach for me, I was sure that C-small was easy and could be coded quickly. (I had and still have no idea how I could successfully approach C-large.)

In particular, I thought that the training budget should always be spent on the worst (lowest-p) cores first. So I wrote this trivial loop, ahem, recursion, that spends as much as possible on those worst cores until either the budget is used up, or the worst cores become improved enough to be as good as the next-better core, or the cores reach p=1. But all my attempts were rejected,  and embarrassingly I still don't know why. Probably my reasoning was again oversimplified. Or it was something about the Lisp representation of floating point numbers. Anyway, the non-functioning code is in C.lisp.

Conclusion

This was the end of my participation in Google Code Jam 2017, because with rank #1850 I didn't make the cut—only the top 1000 in each of the round 1 sub-rounds make it to round 2. Even if I had been successful with the small dataset of C, I wouldn't have made the top 1000. Had I successfully solved the large dataset of B with some time to spare, then it might have worked.

It was mostly fun to code these small problems in Lisp. My Lisp skills are getting a bit rusty though. And I'm still not as comfortable with the SBCL/SLIME combination as I am with Allegro Common Lisp and its Emacs interface, but as I had mentioned in previous posts, I wanted to use completely free/open source software. And the awesome Allegro debugger wouldn't have been of much help here anyway.

And I love the problems! Kudos to the Googlers who come up with them.

How did the other Lispers fare? Badly as far as I can see. There haven't been many Lisp submissions in 2017 anyway, compared with 2016.

Enabling BBR on Ubuntu Xenial

BBR (Bottleneck Bandwidth and RTT) is a novel congestion control scheme for TCP. It was developed by several Google engineers and has been in use on Google.com and Youtube. See the PERT KB for more background.
Here's how you can enable it on a machine running Ubuntu 16.04 LTS "Xenial Xerus".
Note: This is mostly useful for systems that send fairly large TCP flows. And the configuration is incompatible with other packet schedulers such as fq_codel, so consider the tradeoff.

Upgrade Linux Kernel to ≥4.9

The BBR implementation was merged into the master branch of the Linux kernel before the 4.9 release. Xenial uses the 4.4 kernel by default, but you can install a newer one (4.10 as of this writing) from a preview repository of the "hardware enablement" kernel provided by Canonical:

sudo apt install -y --install-recommends \
  linux-generic-hwe-16.04-edge

Enable fq Packet Scheduler

BBR fundamentally requires a packet scheduler that can "pace" outgoing packets per flow.  Enable this by adding an "up" clause (shown in boldface below) to each interface in its corresponding configuration file, e.g. /etc/network/interfaces.d/eth0.cfg:

auto eth0
iface eth0 inet dhcp
  up tc qdisc replace dev $IFACE root fq pacing
iface eth0 inet6 auto

Make BBR the Default Congestion Control Algorithm for TCP

echo net.ipv4.tcp_congestion_control=bbr | \
  sudo tee /etc/sysctl.d/90-bbr

Activate by Rebooting

A reboot should now enable all necessary components: the recent kernel, a pacing-capable packet scheduler, and BBR as the default.

You can check whether BBR is actually in effect by using the ss -i command while having at least one TCP connection open—which is already guaranteed to be the case if you are logged in over SSH.

$ ss --tcp -i
State       Recv-Q Send-Q                                               Local Address:Port                                                                Peer Address:Port
ESTAB       0      1848                         2001:620:5ca1:2f0:f816:3eff:fe9c:71af:ssh                                                           2001:620:0:69::107:57795
  bbr wscale:5,7 rto:208 rtt:6.078/0.237 ato:48 mss:1182 cwnd:12 bytes_acked:156685 bytes_received:10033 segs_out:851 segs_in:977 send 18.7Mbps lastrcv:4 lastack:4 pacing_rate 87.2Mbps unacked:4 retrans:0/13 rcv_rtt:12 rcv_space:28560
ESTAB       0      0                            2001:620:5ca1:2f0:f816:3eff:fe9c:71af:ssh                                                           2001:620:0:69::107:61315
  bbr wscale:5,7 rto:220 rtt:18.674/25.752 ato:40 mss:1182 cwnd:11 bytes_acked:8025 bytes_received:5125 segs_out:86 segs_in:122 send 5.6Mbps lastsnd:470176 lastrcv:480332 lastack:470068 pacing_rate 60.5Mbps rcv_rtt:12 rcv_space:28560

The "-i" part of the output mentions "bbr" and includes some BBR-specific metrics as well as the pacing rate applied to each outbound TCP flow.

Google Code Jam 2017 in Lisp: Qualification Round

In 2016 I was quite happy with my performance in GCJ, even though I didn't make it past Round 1. So despite being quite busy, I registered for the 2017 edition as well. The Qualification Round happened yesterday. It lasted for 27 hours, starting at 1AM in my time zone. I looked at the puzzles after breakfast, and as usual found them intriguing. I initially skipped the first one (Oversized Pancake Flipper) because I knew it must be easy but I didn't find an obvious approach right away. The general approach I took was to first read the problem (out loud if a family member would be willing to listen :-), then think about it while scribbling in a notebook, then when I thought I had an efficient (should be "efficient enough") solution code it up in Lisp more or less top-down. For your amusement, I'm sharing the entirety of my this year's notes, including errors and culs-de-sac.

B. Tidy Numbers

This seemed easy and linear, and I didn't have to do much scribbling. We're looking for the largest Tidy Number that's less than or equal to the input number. We represent the input as a vector of digits. Then we start from the left and look at consecutive pairs of numbers. As long as d_i≤d_i+1 we're happy. If d_i>d_i+1, we need to "fix it up". Fixing it up means usually means changing d_i to d_i-1 and replacing everything to the right with nines. We cannot run into d_i=0, because there are no leading zeroes and we have already checked that the prefix of the number up to and including digit i is Tidy, so the digits cannot become lower. The local fixup may result in the overall number not being Tidy anymore, i.e. we may have made d_i<d_i-1. So we need to recurse (or reiterate, but hey we're thinking in Lisp, right?) on the result. We also may have introduced a leading zero that should be removed. So we remove those. The resulting code is in B.lisp. The code is recursive but not purely functional, because it munches the array using (setf (aref ...) ...). That's ugly and might have made debugging harder, but then the code should be easy enough to not need debugging. And probably this (half-) destructive approach saves a few cycles...

C. Bathroom Stalls

This was definitely the one I had the most (and longest) fun with. It took me several hours (mostly unfocused/doing other things) to get to the breakthrough, which would of course have killed me in a "real" time-limited round.

At first I drew up some simple/small examples. Starting from initial settings with (N=) 7, 6, 5, or 4 spaces, I added all N respective occupants K in order, to get an intuitive picture of how the game plays out. But I didn't find the general pattern.

Pretty quickly, I had a complete understanding of the case where N=2ⁱ-1: The first person (K=1=2⁰) has 2^i-1-1 space to the left and to the right, the next two (2¹) persons each have 2^i-2-1 space to the left and right, and so on. But how to generalize this to Ns that do not have this form?
At any "round of the game" (K), there is a set of "extents" of free space. In each round, the largest of the current extents is "split": If the size of this largest extent is odd and larger than one, that is, of the form 2x+1, then it will be split into two extents of size x. That was the "easy" case that covers the problem completely when N=2ⁱ-1.

But for other kinds of N, I still had no good idea of the pattern how the overall set of extents develops. In a desperate attempt, I tried to draw this up as a tree for N=6. That looked somewhat promising, but I still couldn't figure out the general pattern. At least it shows the special case of the size-2 extent, which doesn't really get split into two extents but degenerates into a single size-1 extent. Hm, this seems complicated.

In an even more desperate attempt, I decided that I needed to look at a bigger example and started drawing up the tree for N=230. Nothing very particular about that number, except that I wanted to start with an even number because that seems to be the more complicated case.
Now that led to the breakthrough. In this non-trivial example, I saw that each level of the tree only contained (a maximum of) two different values with a difference of one.
So a complete and useful description of a given level of the tree is the distribution between these two different values. This is what I wrote up in the right-hand column: On level 1 (after the first split) we have extents of 1x114 and 1x115, on level 2, 1x56 and 3x57, on level 3, 1x27 and 7x28, on level 4, 9x13 and 7x14, and so on. Generalizing to level 0, we can say that we start with 1x230 (1xN) and 0x231.
But how does one get from level j to level j+1? Well, each odd extent splits into two equal extents, and each even one splits into one that's half the original number and one that's one lower. That leads to the two cases below the large (partially written-out) tree. (I started writing the cases with m and n and then replaced that with a and b, respectively, but didn't do so all the way because I had enough to write the code.) The code is in C.lisp, and I'm quite happy with it. Initially I had switched the even and odd cases, which led to wrong results on the tiny example under the problem description. First I panicked, but by simply running a few examples under (trace minmax-1) I could locate the error. I knew the code was O(log(N)) efficient, so I fed it the small, still-small, and large cases with confidence, and each was processed in less than 10 milliseconds.

A. Oversized Pancake Flipper

At this point I was fairly certain to have passed the mark with just two problems done, assuming that at least one of my "large" data set solutions were correct. But it nagged me that I hadn't turned in A, which was supposedly the easiest puzzle. I thought about it again and came up with the following reasoning: The basic operator (multi-pancake flip) is commutative—it doesn't matter whether I flip 4–6 and then 3–5, or the other way round. That means we should be able to strictly process the pancakes in some order, e.g. from left to right. The operator is also an inversion... applying it twice returns the original state and, because of the commutative property, never makes sense.

So we can just go to the row of pancakes left-to-right. If the current pancake is –, we turn it and the others to its right. If the current pancake is + we leave it. Then we shift one place to the right and repeat. When we're at the end of the row (which is K places before the end because of the flipper restriction), we're done. The last pancakes must then all be + or we lose.

The code is in A.lisp. Again, this is written recursively but destructively: The flipping is done by modifying elements in the vector (string). Probably the code would be more elegant if rewritten using a list instead of a vector, and constructing a fresh list when flipping.

When reading the initial state of pancakes, I skipped the reverse (or nreverse) step because it really doesn't matter whether we operate on a mirror image.

D. Fashion Show

Wow, that problem sure seemed from a different league. Fewer than 1'000 contestants even got the small solution right, whereas all other problems had at least 13'000 correct solutions. My mental analysis and scribbling didn't get me very far. You cannot have more than one non-+ in the same row or column, not more than one non-x in the same diagonal. If you want to have an o (twice the value of a + or x), then all the other non-empty models in the same row/column need to be +, all on the same diagonal need to be x.

There's also a parallel to the N-queens problem, in that in both cases you cannot have more than one "thing" (where thing means "queen" or "non-+/x") per row/column/diagonal. But where N-queens only has one "thing", the queen, in this case the things are different (non-+/non-x), and what's more we need to maximize some value.

So I couldn't figure out a useful approach, and the fact that some initial models have been placed makes the reasoning even harder. So I gave up before even trying very hard.

Conclusion

The next morning I checked my results for the large data sets and found them all correct. I ranked in the 4'400s, but then I started late. My 65 points would have been sufficient to enter the top 1000 if I had finished everything in 2h15. That would have been somewhat unlikely, even if I had started immediately and had been fully focused. The first sub-round of round 1 will start at 3AM local time, so I'll probably skip it and try 1B a week later, and then probably 1C. Let's see!

I didn't have too much trouble with the SLIME/SBCL combination; for more than 25 years I have mostly been using Allegro and its Emacs interface when programming, and this is still much easier for me, especially when debugging. But for the contest I really wanted to use a Free/Open Source Lisp.

Google Code Jam 2016 in Lisp: Round 1B/1C

After failing to qualify in round 1A in this year's Google Code Jam, I had two more chances to advance to round 2.

Round 1B

In round 1B, I only managed to solve the first problem, Getting The Digits.  It took me quite some time, but eventually I found the recommended approach to count letters that are specific to a single digit (e.g. Z occurs in no other digits than ZERO), then eliminate those digits; initially all the even digits have unique characters (0:Z, 2:W, 4:U, 6:X, 8:G), but after those have been eliminated, this holds for other digits as well: (1:O, 3:T, 5:F, 7:S) and finally (9:I).  The program is ugly in that this sequence of letter/digit combinations is hard-coded; a good program would extract the letters by itself so that it could be ported more easily to other languages etc.  My Lisp code is here.

Anyway, both of the other problems resisted my analysis.  I should have been able to solve Close Match, but I couldn't decide whether it should be solved from left to right or right to left, and ended up spinning in circles.  The final problem, Technobabble, is beyond my current capabilities: I neither succeeded in mapping the problem description to the correct graph problem (minimum edge cover on a bipartite graph), nor would I have been able to code an efficient solution to that problem.  As an undergraduate I was good at these kinds of things, but lacked practice over the past 20 years or so.

With only one problem set solved, I ranked #4259 and was very far from qualifying.  The other Lispers didn't fare much better in this round, except for orivej, who advanced by ranking #847.  He or she only used Lisp for the first problem, and Python for the third.

Round 1C

Again I was able to solve the first (and easiest) problem, Senate Evacuation.  There are many viable approaches, and it took me a bit too long to find one.  Here's my slightly embellished A.lisp (raw submission).

After 35 minutes, I attacked the second problem, Slides! After lots of scribbling of graphs and diagonal matrices, I found a method to construct slides for an arbitrary M below the maximum that maps nicely to binary numbers.  The resulting code (B.lisp, cleaned up from the actual submission) is short and sweet and even turns out to work.  But again it has taken me much too long to get there.

With only 16 minutes left, I looked at the final problem, Fashion Police.  I knew I didn't have the slightest chance to solve the problem properly, so I started coding a brute-force algorithm that might solve the small dataset.  But that was not as small as it looked, and the problem not as simple as I thought, so I failed.  Well, anything else would have been a miracle.

My final rank was in this round #1326.  I didn't qualify, but I was quite happy with this honorable (for me) result.  To qualify, I would have had to finished my two working programs in 1h39'30" instead of 2h13'46".  Maybe with more practice I'd be able to save those 34 minutes somewhere—but it looks like a long shot!

Looking at the other Lispers in the contest: This round saw an awesome result from lpetru, who handed in a perfect score (100 points, all problem sets solved), landing him or her on rank #98.  And so far he or she submitted all solutions in Lisp.  I know whom to root for in round 2!

Final remarks

Again it was lots of fun to participate in Google Code Jam.  It was exhausting and sometimes frustrating (especially in round 1B).  But it's good to have those old brain cells work outside their comfort zone.  So I'll probably try again and hope that I'll eventually manage to reach round 2.  See you next year!

Google Code Jam 2016 in Lisp: Round 1A

Having survived (like 96% of all Lispers and 81.5% of the wider population) the Qualification Round, I'm trying to pass the first "real" round. As expected, things are getting much more serious now! Round 1 is held in three sub-rounds, and you can participate in each, or until you advance to round 2.  In each sub-round, the 1'000 highest-ranking contestants advance to round 2.

Where the qualification round had consisted of four problems (subjectively, two easy and two somewhat tricky) with 27 hours time to solve them, round 1 is three problems (an easier one and two quite tricky ones) that have to be solved within 2h30.  From previous participations I knew that I would likely run into the time limit.

Getting off to a good start

Although the first sub-round, Round 1A, started at 3AM in my time zone, I decided to give it a try.  I went to bed somewhat early and asked one of my sons, who was going out, to wake me up a bit before three, which he dutifully did.  (Thank you!)

The first puzzle (The Last Word) went quite well.  Somehow I convinced myself quickly that it is sufficient to put the next letter in front if it is greater than or equal to the first letter of the current substring, and at the end if it is smaller.  My straightforward Lisp implementation can be found here.

The second task (Rank and File) was cute.  After a bit of scribbling and thinking, I suddenly realized that it's sufficient to look for numbers that occur an odd number of times, sort those numbers and be done with it.  I coded this up fairly quickly as well (Lisp source here).

So I was done with the first two problems after 44'22" with no penalties.  I was also reasonably certain that I got both the large input sets right as well.  (In the Qualification Round I had wrongly believed so as well, but in the critical case of problem C, the difference between small and large input was qualitative rather than quantitative.)

Problem C: close, but no cigar

With 1h45 left, I attacked the last problem, BFFs.  This is a great problem—very simple in its formulation, yet requiring good (but not impossible) analysis, and also a bit of programming skills.  It would be a nice recruitment interview problem.

This is pretty clearly a graph problem.  The input set is a rather special set of directed graph with out-degree 1 (the BFF function).  This must somehow be mapped to another kind of graph that represents a circle with some constraints.  But this mapping is not that straightforward.  Studying the sample inputs on the problem page, I quickly found that there are two different types of circle: One is a complete cyclic sub-graph of the input—that's easy to find given the simple type of input graph.  The other possibility is to have a "core" circular graph of two nodes (a "couple" of mutual BFFs), with "extensions" of other nodes (kids) that (BFF-) point towards the nodes of the core.  I wrote the code for both cases and tested it successfully against the tiny example on the problem page.  My code looked for the maximum cycle, and the maximum graph of the second type, and would return the size of the bigger one.

To my great disappointment, the system told me that my solution for the small input set was incorrect.  I couldn't figure out what was wrong, and time ran out on me.  It was 5h30, and I had finished at rank #2137, with correct solutions for the first two problems and none for the third.

I went to bed slightly frustrated, and of course couldn't sleep right away but continued to think about the problem... what if the entire class consists of happy couples, i.e. mutual BFFs? Then I could put everybody in the circle.  Hmmm... that's a case I had somehow missed! In fact the second type of graph (couple with "inbound" extensions) can occur arbitrarily often in the circle.  Having understood my error, I could finally sleep.

In the next morning I had to write the code, although the contest was over.  Although I had coded (correctly, I think) a good part of the necessary logic the night before, it took me several hours to get the algorithm right for the actual solution.  I was too ambitious and wanted to write a really efficient solution ("early optimization..."), although a stupid O(N²) approach would probably have worked even for the maximum problem size of a 1'000 kids class.  Probably I was also tired.  Also, I saw that I don't know how to do proper debugging in SBCL, and had written the code in a way that made debugging somewhat hard.  Finally I got the code right, see C.lisp here.  Although the code is not particularly compact, I'm happy with it because at least I can read and understand it (for how long is another question), and it is O(N) with a pretty low constant factor for what that is worth—it runs in 20 milliseconds on the large practice set including I/O.

What about the other Lispers?

The per-language dashboard for Lisp on go-hero.net shows that none of us advanced in sub-round 1A.  A very honorable mention goes to vii, who ended up on rank #1019 with correct solutions for all problems and input sets, but with a few penalties landing him or her 2'33" above the qualifying threshold.  Good job & better luck next time! I'm sure he or she has what it takes.

Will I continue to use Lisp (and SBCL)?

While coding the last problem (BFFs), I sometimes thought that the program would have been easier to write in another language, even in C.  I couldn't find an elegant recursive approach, and I always found writing iterative code in Lisp a bit awkward; I never really mastered loop or any of the other "modern" iteration constructs—why learn them when do will do? But the primary reason why I would have preferred another language was data structures.  I built my graph from "node" objects (expressed as defstructs) with pointers, which was natural enough and worked well.  But in another language I would have used a convenient vector representation... in Lisp the syntax for dereferencing vectors is so verbose and awkward that I was worried getting things wrong.  And I felt that in the short time, I couldn't write a nice data type abstraction to hide the intricacies of such a vector representation.

What about SBCL? As mentioned before, I normally use Franz Inc's excellent Allegro Common Lisp, but I wanted to use a "real" open source system for the contest.  I'm slowly getting used to SLIME after more than 20 years of Franz's Emacs-Lisp interface.  But when I had to debug my final BFFs solution, I felt almost blind because I wasn't able to look at local variables from within a backtrace, something that works quite nicely with Allegro.  So I had to resort to the equivalent of printf()s (warn), which made me feel a bit dirty.  Probably I should learn how others debug under SBCL, I'm sure there's a better way! [Edited to add: (declaim (optimize (debug 3))) seems to go a long way!]

In the weeks before the contest, I had toyed with the idea to use Go this year, but I didn't have the time to learn the language well enough to make this viable.  And overall, Lisp has treated me quite well this year.  Also I've been using it for about 30 years now.  So I don't feel like quitting mid-contest.

There are still two opportunities for me to try to advance to round 2.  Timezone-wise they are much more convenient for me.  On the other hand, I need some luck to find a good solution for the medium-tough problem as quickly as in 1A, and a bit more to find a good solution for the tough problem at all.  But I'm an eternal optimist...

Google Code Jam 2016 Qualification Round in Lisp

This year I registered for Google Code Jam again. The qualification round was this weekend, and I submitted solutions in Lisp.  This time I used SBCL with GNU Emacs and SLIME as the IDE, rather than Allegro Common Lisp with GNU Emacs and Franz Inc's emacs-lisp interface.
The puzzles were lots of fun, as usual.  A (Counting Sheep) and B (Revenge of the Pancakes) were easy.  When I looked at C (Coin Jam), I found it tricky, so I put that aside and did D (Fractiles) next.  Explaining the problem to my 19-year-old son helped me find a good approach.  After implementing it and submitting the solutions, I turned back to puzzle C.  My solution looked complex, and I suspected that there would be a more elegant way.  Still I was happy with the solution as it produced a result for the large set very quickly, so I submitted the results and was done. My solutions can be found here.

So how did I fare?

The round went on for 27 hours, and I started somewhere in the middle due to timezone.  So after I was done with my solutions, I had to wait until the next day to see the final score.  The "small" problem datasets are checked immediately after submission, so I already knew I had all of those right before I went to bed.  Since they already accounted for 37 points, and I needed 40 to advance to the next round, I could be pretty optimistic that I had passed—having any large dataset correct would get me more than the three missing points.

The intermediate scoreboard put me on rank #1189; this is assuming (for everybody) that all the "large" solutions that were submitted would be correct.  I was pretty sure that mine were all correct, so I had hopes that my final rank would be even higher.

But my final rank was #2748, because it turned out that my "large" solution for D (Fractiles) was incorrect.  This was a big disappointment, because I was very confident in my solution approach.  And it turned out that my error was a stupid one in one of the "easy" parts of the implementation; the decision whether the number of grad students (S) is sufficient to clean enough tiles in the base-length-K and complexity C artwork to detect whether there's any gold tile.  The correct test would be

  (if (< (* s c) k)
      "IMPOSSIBLE"

...but I used:

  (let ((min-s (truncate k c)))
    (if (< s min-s)
        "IMPOSSIBLE"

My test was too lax, and I sometimes output a "solution" when I should have printed IMPOSSIBLE.  This stupid mistake cost me >2000 ranks.  It would be nice if I managed to avoid this in the next round.  But realistically I'll make even more of these mistakes because the problems will be harder and the time pressure much higher.

How did the other Lispers do?

On www.go-hero.net there are statistics about GCJ participants and submissions, and it's easy to find all submissions for this round that used Lisp.  We Lispers were quite successful this weekend, with 26 out of 27 proceeding to the next round, a success rate of 96% compared to 81.5% for the entire population.
The highest ranking Lisper was Ipetru at #594 with all solutions in Lisp (and of course all correct).  I looked at his solutions for C and D, and they are so incredibly compact that I couldn't believe my eyes.  D used the same approach as I had, just very elegantly written—the code proper is about two lines; much harder to hide stupid mistakes in there! C used a completely different approach, deterministically generating Jamcoins rather than checking candidates as I had.
The second-ranking Lisper was DarkKnight. at #720.  He only wrote one solution in Lisp.  In fact he used different languages for all solutions, and I mean all 8 solutions, not all 4 puzzles! bc, OCaml, Lisp, Racket, Lua, Perl, Octave, R.  Impressive! :-)