Re: Hydra doesn't have enough memory+swap to evaluate jobsets

[Ludovic Courtès]

(This is the continuation of a thread about the memory requirements of
job evaluation on Hydra.)

Here’s a run in conditions similar to those of Hydra:

--8<---------------cut here---------------start------------->8---
$ \time -v guile -l build-aux/hydra/gnu-system.scm -c '(use-modules (guix)) 
(with-store store (hydra-jobs store (list)))'
prepending "build-aux/hydra/../.." to the load path
;;; note: source file build-aux/hydra/../../guix/config.scm
;;;       newer than compiled 
/home/ludo/soft/share/guile/site/2.0/guix/config.go
;;; note: auto-compilation is enabled, set GUILE_AUTO_COMPILE=0
;;;       or pass the --no-auto-compile argument to disable.
;;; compiling build-aux/hydra/../../guix/config.scm
;;; compiled 
/home/ludo/.cache/guile/ccache/2.0-LE-8-2.0/home/ludo/src/guix/guix/config.scm.go
;;; note: source file build-aux/hydra/../../guix/store.scm
;;;       newer than compiled /home/ludo/soft/share/guile/site/2.0/guix/store.go
;;; compiling build-aux/hydra/../../guix/store.scm
;;; note: source file build-aux/hydra/../../guix/utils.scm
;;;       newer than compiled /home/ludo/soft/share/guile/site/2.0/guix/utils.go
;;; compiling build-aux/hydra/../../guix/utils.scm

[...]

;;; compiling build-aux/hydra/../../gnu/packages/wdiff.scm
;;; compiled 
/home/ludo/.cache/guile/ccache/2.0-LE-8-2.0/home/ludo/src/guix/gnu/packages/wdiff.scm.go
;;; compiling build-aux/hydra/../../gnu/packages/gps.scm
;;; compiled 
/home/ludo/.cache/guile/ccache/2.0-LE-8-2.0/home/ludo/src/guix/gnu/packages/gps.scm.go
        Command being timed: "guile -l build-aux/hydra/gnu-system.scm -c 
(use-modules (guix)) (with-store store (hydra-jobs store (list)))"
        User time (seconds): 90.48
        System time (seconds): 0.30
        Percent of CPU this job got: 103%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 1:27.40
        Average shared text size (kbytes): 0
        Average unshared data size (kbytes): 0
        Average stack size (kbytes): 0
        Average total size (kbytes): 0
        Maximum resident set size (kbytes): 777056
        Average resident set size (kbytes): 0
        Major (requiring I/O) page faults: 3
        Minor (reclaiming a frame) page faults: 18296
        Voluntary context switches: 7426
        Involuntary context switches: 26029
        Swaps: 0
        File system inputs: 7392
        File system outputs: 11936
        Socket messages sent: 0
        Socket messages received: 0
        Signals delivered: 0
        Page size (bytes): 4096
        Exit status: 0
--8<---------------cut here---------------end--------------->8---

So there’s a peak at 758 MiB, which is arguably too much, but far from
5 GiB.

Each evaluation forces a complete recompilation.  However this is
apparently not what takes the most memory:

--8<---------------cut here---------------start------------->8---
$ \time -v ./pre-inst-env guile -l build-aux/hydra/gnu-system.scm -c 
'(use-modules (guix)) (with-store store (hydra-jobs store (list)))'
;;; note: auto-compilation is enabled, set GUILE_AUTO_COMPILE=0
;;;       or pass the --no-auto-compile argument to disable.
;;; compiling /home/ludo/src/guix/build-aux/hydra/gnu-system.scm
prepending "build-aux/hydra/../.." to the load path
;;; compiled 
/home/ludo/.cache/guile/ccache/2.0-LE-8-2.0/home/ludo/src/guix/build-aux/hydra/gnu-system.scm.go
prepending "build-aux/hydra/../.." to the load path
        Command being timed: "./pre-inst-env guile -l 
build-aux/hydra/gnu-system.scm -c (use-modules (guix)) (with-store store 
(hydra-jobs store (list)))"
        User time (seconds): 45.26
        System time (seconds): 0.14
        Percent of CPU this job got: 104%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:43.48
        Average shared text size (kbytes): 0
        Average unshared data size (kbytes): 0
        Average stack size (kbytes): 0
        Average total size (kbytes): 0
        Maximum resident set size (kbytes): 798912
--8<---------------cut here---------------end--------------->8---

The latter is faster, but uses about as much memory as the
auto-compilation variant.

For reference, loading all the package modules takes around 180 MiB:

--8<---------------cut here---------------start------------->8---
$ \time -v guix build foobarbaz
guix build: error: foobarbaz: unknown package
Command exited with non-zero status 1
        Command being timed: "guix build foobarbaz"
        User time (seconds): 0.31
        System time (seconds): 0.03
        Percent of CPU this job got: 64%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.53
        Average shared text size (kbytes): 0
        Average unshared data size (kbytes): 0
        Average stack size (kbytes): 0
        Average total size (kbytes): 0
        Maximum resident set size (kbytes): 181376
--8<---------------cut here---------------end--------------->8---

Anyway, we need to do actual heap profiling now.  :-)

Ludo’.