Skip to content

Make MediaWiki MLP jobs tunable + add reportable tuned preset (#702)#702

Open
charles-typ wants to merge 1 commit into
facebookresearch:v2-betafrom
charles-typ:export-D110121020-to-v2-beta
Open

Make MediaWiki MLP jobs tunable + add reportable tuned preset (#702)#702
charles-typ wants to merge 1 commit into
facebookresearch:v2-betafrom
charles-typ:export-D110121020-to-v2-beta

Conversation

@charles-typ

@charles-typ charles-typ commented Jun 30, 2026

Copy link
Copy Markdown
Contributor

Summary:

Problem: MediaWiki tuning sweeps on Phoenix (NUMA-1) appeared to produce
"no valid result" in benchpress. The runs themselves were fine -- they
produced metrics JSON + mpstat artifacts -- but they were launched under
custom job names (mw_mlp_r1_tuned, mw_mlp_r1_c512, mw_mlp_r3_tuned) via a
local -j jobs file. benchpress report <name> returns "No job found" for
unregistered names, so those results never rolled into the DCPerf geomean.
From the reporting tool's view, the data did not exist.

Why: The registered oss_performance_mediawiki_mlp* jobs only template
scale_out (-R) and client_threads (-c). To vary server_threads (-t) or
memcache_threads (-m) -- needed for the R1/t400 single-large-HHVM variant --
users had to fork the job under a new name, orphaning the result from
report/geomean.

Fix:

  • Add -t{server_threads} and -m{memcache_threads} to the registered
    oss_performance_mediawiki_mlp, _mlp_no_jit, and _mem jobs. New vars
    default to 0, which run.sh already interprets as "use computed default",
    so stock behavior is unchanged (verified: stock _mlp still expands to
    R2/t200/c256 on a 128-core host).
  • Add a new registered preset oss_performance_mediawiki_mlp_tuned
    (R1/t400/c256/m64) capturing the best-performing tuning variant, so it
    reports + rolls into the geomean like any first-class job.
  • run.sh: guard -m0 to fall back to the computed default, matching the
    existing -R0/-t0/-c0 guards.

Differential Revision: D110121020

@meta-cla meta-cla Bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Jun 30, 2026
@meta-codesync

meta-codesync Bot commented Jun 30, 2026

Copy link
Copy Markdown

@charles-typ has exported this pull request. If you are a Meta employee, you can view the originating Diff in D110121020.

…okresearch#702)

Summary:

Problem: MediaWiki tuning sweeps on Phoenix (NUMA-1) appeared to produce
"no valid result" in benchpress. The runs themselves were fine -- they
produced metrics JSON + mpstat artifacts -- but they were launched under
custom job names (mw_mlp_r1_tuned, mw_mlp_r1_c512, mw_mlp_r3_tuned) via a
local -j jobs file. `benchpress report <name>` returns "No job found" for
unregistered names, so those results never rolled into the DCPerf geomean.
From the reporting tool's view, the data did not exist.

Why: The registered oss_performance_mediawiki_mlp* jobs only template
scale_out (-R) and client_threads (-c). To vary server_threads (-t) or
memcache_threads (-m) -- needed for the R1/t400 single-large-HHVM variant --
users had to fork the job under a new name, orphaning the result from
report/geomean.

Fix:
- Add -t{server_threads} and -m{memcache_threads} to the registered
  oss_performance_mediawiki_mlp, _mlp_no_jit, and _mem jobs. New vars
  default to 0, which run.sh already interprets as "use computed default",
  so stock behavior is unchanged (verified: stock _mlp still expands to
  R2/t200/c256 on a 128-core host).
- Add a new registered preset oss_performance_mediawiki_mlp_tuned
  (R1/t400/c256/m64) capturing the best-performing tuning variant, so it
  reports + rolls into the geomean like any first-class job.
- run.sh: guard -m0 to fall back to the computed default, matching the
  existing -R0/-t0/-c0 guards.

Differential Revision: D110121020
@meta-codesync meta-codesync Bot changed the title Make MediaWiki MLP jobs tunable + add reportable tuned preset Make MediaWiki MLP jobs tunable + add reportable tuned preset (#702) Jun 30, 2026
@charles-typ charles-typ force-pushed the export-D110121020-to-v2-beta branch from 981bd42 to c17182d Compare June 30, 2026 06:47
meta-codesync Bot pushed a commit that referenced this pull request Jun 30, 2026
Summary:
Pull Request resolved: #702

Problem: MediaWiki tuning sweeps on Phoenix (NUMA-1) appeared to produce
"no valid result" in benchpress. The runs themselves were fine -- they
produced metrics JSON + mpstat artifacts -- but they were launched under
custom job names (mw_mlp_r1_tuned, mw_mlp_r1_c512, mw_mlp_r3_tuned) via a
local -j jobs file. `benchpress report <name>` returns "No job found" for
unregistered names, so those results never rolled into the DCPerf geomean.
From the reporting tool's view, the data did not exist.

Why: The registered oss_performance_mediawiki_mlp* jobs only template
scale_out (-R) and client_threads (-c). To vary server_threads (-t) or
memcache_threads (-m) -- needed for the R1/t400 single-large-HHVM variant --
users had to fork the job under a new name, orphaning the result from
report/geomean.

Fix:
- Add -t{server_threads} and -m{memcache_threads} to the registered
  oss_performance_mediawiki_mlp, _mlp_no_jit, and _mem jobs. New vars
  default to 0, which run.sh already interprets as "use computed default",
  so stock behavior is unchanged (verified: stock _mlp still expands to
  R2/t200/c256 on a 128-core host).
- Add a new registered preset oss_performance_mediawiki_mlp_tuned
  (R1/t400/c256/m64) capturing the best-performing tuning variant, so it
  reports + rolls into the geomean like any first-class job.
- run.sh: guard -m0 to fall back to the computed default, matching the
  existing -R0/-t0/-c0 guards.

Reviewed By: YifanYuan3

Differential Revision: D110121020

fbshipit-source-id: 76375b6cc79ffcbea38b159a34a8380da0b3ac59
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. meta-exported

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant