diff options
| author | Maxim Kuvyrkov <maxim.kuvyrkov@linaro.org> | 2018-12-03 10:41:36 +0000 |
|---|---|---|
| committer | Maxim Kuvyrkov <maxim.kuvyrkov@linaro.org> | 2018-12-03 10:41:36 +0000 |
| commit | 036b7091d719f417f1822e870c14e000fb6a432e (patch) | |
| tree | 308c4b13c9c7a1c7893a3bdf475427f36826abdb | |
| parent | 64d2d047f651b25373d53ed6c69b95896aa3590c (diff) | |
tcwg-llvmbot: Limit ninja on avg-load on silent bots, but not on normal bots
There are two sources of massive parallelism in LLVM bots: (1) ninja-driven
builds of LLVM and (2) llvm-lit-driven testsuites.
We are, currently, limiting parallelism under high load for (1), but
can't do it effectively for (2). This causes a situation when "normal"
bots gracefully yield CPU time by running single-threaded ninja to
"silent" bots during their llvm-lit testsuites.
Since the linaro-armv8-01 machine seems to be OK with running under
500+ average process load, this patch removes avg-load limit from "normal"
bots, which will increase the overall system load, but will shift CPU time
towards "normal" bots.
Change-Id: I2dcd27b9e9b800344a1acea0ec17681d5b97c341
| -rwxr-xr-x | tcwg-base/tcwg-llvmbot/run.sh | 24 |
1 files changed, 18 insertions, 6 deletions
diff --git a/tcwg-base/tcwg-llvmbot/run.sh b/tcwg-base/tcwg-llvmbot/run.sh index f9d74916..d3308d49 100755 --- a/tcwg-base/tcwg-llvmbot/run.sh +++ b/tcwg-base/tcwg-llvmbot/run.sh @@ -137,14 +137,26 @@ exec /usr/bin/ninja -j$n_cores "\$@" EOF ;; *) - # Throttle ninja on system load, system memory and container memory limit. - # When running with "-l 2*N_CORES -m 50 -M 50" ninja will not start new jobs - # if system or container memory utilization is beyond 50% or when load is - # above double the core count. Ninja will also stall up to 5 seconds (-D 5000) - # before starting a new job to avoid rapid increase of resource usage. + # Throttle ninja on system load, system memory and container memory + # limits. + case "$1" in + lab.llvm.org:9994) + # Run silent bots with single-threaded ninja when average load + # is beyond twice the number of cores. + avg_load_opt="-l $((2*$n_cores))" + ;; + *) + avg_load_opt="" + ;; + esac + # Make ninja run single-threaded if system or container memory + # utilization is beyond 50% (-m 50 -M 50). + # Make ninja stall for up to 5 seconds (-D 5000) before starting + # a new job when usage decreases under threshold (to avoid rapid + # increase of resource usage from N_CORES-1 new processes). cat > /usr/local/bin/ninja <<EOF #!/bin/sh -exec /usr/local/bin/ninja.bin -j$n_cores -l $((2*$n_cores)) -m 50 -M 50 -D 5000 "\$@" +exec /usr/local/bin/ninja.bin -j$n_cores $avg_load_opt -m 50 -M 50 -D 5000 "\$@" EOF ;; esac |