[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

When substitute download + decompression is CPU-bound

From: Ludovic Courtès
Subject: When substitute download + decompression is CPU-bound
Date: Mon, 14 Dec 2020 23:20:17 +0100
User-agent: Gnus/5.13 (Gnus v5.13) Emacs/27.1 (gnu/linux)

Hi Guix!

Consider these two files:

Quick decompression bench:

--8<---------------cut here---------------start------------->8---
$ du -h /tmp/uc.nar.[gl]z
103M    /tmp/uc.nar.gz
71M     /tmp/uc.nar.lz
$ gunzip -c < /tmp/uc.nar.gz| wc -c
$ time lzip -d </tmp/uc.nar.lz >/dev/null

real    0m6.040s
user    0m5.950s
sys     0m0.036s
$ time gunzip -c < /tmp/uc.nar.gz >/dev/null

real    0m2.009s
user    0m1.977s
sys     0m0.032s
--8<---------------cut here---------------end--------------->8---

The decompression throughput (compressed bytes read in the first column,
uncompressed bytes written in the second column) is:

          input   |  output
  gzip: 167 MiB/s | 52 MB/s
  lzip:  56 MiB/s | 11 MB/s

Indeed, if you run this from a computer on your LAN:

  wget -O - … | gunzip > /dev/null

you’ll find that wget caps at 50 M/s with gunzip, whereas with lunzip it
caps at 11 MB/s.

>From my place I get a peak download bandwidth of 30+ MB/s from, thus substitute downloads are CPU-bound (I can’t go
beyond 11 M/s due to decompression).  I must say it never occurred to me
it could be the case when we introduced lzip substitutes.

I’d get faster substitute downloads with gzip (I would download more but
the time-to-disk would be smaller.)  Specifically, download +
decompression of ungoogled-chromium from the LAN completes in 2.4s for
gzip vs. 7.1s for lzip.  On a low-end ARMv7 device, also on the LAN, I
get 32s (gzip) vs. 53s (lzip).

Where to go from here?  Several options:

  0. Lzip decompression speed increases with compression ratio, but
     we’re already using ‘--best’ on ci.  The only way we could gain is
     by using “multi-member archives” and then parallel decompression as
     done in plzip, but that’s probably not supported in lzlib.  So
     we’re probably stuck here.

  1. Since still provides both gzip and lzip archives,
     ‘guix substitute’ could automatically pick one or the other
     depending on the CPU and bandwidth.  Perhaps a simple trick would
     be to check the user/wall-clock time ratio and switch to gzip for
     subsequent downloads if that ratio is close to one.  How well would
     that work?

  2. Use Zstd like all the cool kids since it seems to have a much
     higher decompression speed: <>.
     630 MB/s on ungoogled-chromium on my laptop.  Woow.

  3. Allow for parallel downloads (really: parallel decompression) as
     Julien did in <>.

My preference would be #2, #1, and #3, in this order.  #2 is great but
it’s quite a bit of work, whereas #1 could be deployed quickly.  I’m not
fond of #3 because it just papers over the underlying issue and could be
counterproductive if the number of jobs is wrong.



reply via email to

[Prev in Thread] Current Thread [Next in Thread]