Make prod-aligned knobs the defaults (#712) - #712

Closed

excelle08 wants to merge 38 commits into

facebookresearch:v2-betafrom

excelle08:export-D107579495-to-v2-beta

Commits on Jul 3, 2026

Add scaled-up feature extraction pipeline

excelle08
authored and
meta-codesync[bot]
committed
07d7197
Refactor feature extractors to match production instruction mix

excelle08
authored and
meta-codesync[bot]
committed
2ad0723
Replace oldisim framework with folly-based FeedSimServer/FeedSimDriver

excelle08
authored and
meta-codesync[bot]
committed
2d162f0
Add Silesia corpus client-server story pipeline

excelle08
authored and
meta-codesync[bot]
committed
a8aea32
Match production request size distribution in DriverNodeRank

excelle08
authored and
meta-codesync[bot]
committed
bee936d
Refactor LeafNodeRank: split DLRM functions, async DLRM, Silesia response generator

excelle08
authored and
meta-codesync[bot]
committed
6b715a1
Name thread pools to match production: ThriftSrv.IO, SREventBase, RANKER, GlobalCPUThread

excelle08
authored and
meta-codesync[bot]
committed
4a34629
Add mock_services standalone Thrift server (20 methods, real fbthrift)

excelle08
authored and
meta-codesync[bot]
committed
c614da5
Add 5 prod-shaped thrift structs and register new request type IDs

excelle08
authored and
meta-codesync[bot]
committed
cf18c8e
Migrate compression to ManagedCompression

excelle08
authored and
meta-codesync[bot]
committed
04fef7e
Wire LeafNodeRank to mock_services for outbound RPC fanout

excelle08
authored and
meta-codesync[bot]
committed
2ecde9f
Add SemiFuture driver API + RunSession + first-story latency

excelle08
authored and
meta-codesync[bot]
committed
5698c53
Replace shim handlers with 5 real per-method server handlers

excelle08
authored and
meta-codesync[bot]
committed
839543e
Add LatencyHistogram instrumentation for mock_services fanout

excelle08
authored and
meta-codesync[bot]
committed
bec2c83
Add mock_services latency shaping knobs (cap/offset/skip-threshold)

excelle08
authored and
meta-codesync[bot]
committed
0a10a55
Ship rpc_dist_v2.json and add p99_9/p99_99 percentile support

excelle08
authored and
meta-codesync[bot]
committed
66fb4d3
Drop max(1, ...) floor in issueOutboundFanout

excelle08
authored and
meta-codesync[bot]
committed
a05298f
Remove srvIOThreadPool and its throw-away datagen + compression

excelle08
authored and
meta-codesync[bot]
committed
b815f10
Add RPC_FANOUT_SCALE env override in run.sh

excelle08
authored and
meta-codesync[bot]
committed
5bee519
Move runFeatureExtraction off the dispatcher; enable for dlrm_mini

excelle08
authored and
meta-codesync[bot]
committed
b952a9c
Tune feedsim_autoscale_dlrm_mini toward production CPU profile (1800 stories, 8 inferences)

excelle08
authored and
meta-codesync[bot]
committed
ffd3011
Raise default --io_threads (ThriftSrv.IO pool) from 4 to nproc

excelle08
authored and
meta-codesync[bot]
committed
91f39bf
Bound issueOutboundFanout per-request concurrency with folly::window(K=32)

excelle08
authored and
meta-codesync[bot]
committed
035991a
DriverNodeRank: per-second QPS trace + soft in-flight cap

excelle08
authored and
meta-codesync[bot]
committed
15423f8
LeafNodeRank: kOutboundFanoutWindow=16, driver_threads=nproc/4 on SMT-on

excelle08
authored and
meta-codesync[bot]
committed
6473424
Cap PyTorch thread pools to avoid nproc^2 GlobalCPUThread explosion

excelle08
authored and
meta-codesync[bot]
committed
f3ef32a
Run one mock_services per feedsim instance, taskset-isolated

excelle08
authored and
meta-codesync[bot]
committed
50ea249
Wire breakdown.csv into feedsim_autoscale_dlrm + add preprocessing/postprocessing logs

excelle08
authored and
meta-codesync[bot]
committed
823405f
Round-robin outbound RPCs across one MockServicesClient per SREventBase

excelle08
authored and
meta-codesync[bot]
committed
b873cc4
Channel keepalive eliminates cold-channel latency cliff

excelle08
authored and
meta-codesync[bot]
committed
d1c6354
Fix concurrent SIGSEGV in DLRM forward() and feature extractors

excelle08
authored and
meta-codesync[bot]
committed
0939d5f
Match prod multifeed RPC profile: ZSTD, smaller response, DLRM=1, feature extractors

excelle08
authored and
meta-codesync[bot]
committed
f972432
Fix perf hook teardown hang via process-group signaling and bounded waits

excelle08
authored and
meta-codesync[bot]
committed
6df773c
Add start_new_session=True to mpstat Popen to fix perf hook teardown hang

excelle08
authored and
meta-codesync[bot]
committed
176c657
Enable TLS-over-Rocket on mock_services with ALPN rs

excelle08
authored and
meta-codesync[bot]
committed
bbe8d9a
Pre-build 64-slot response template pool to eliminate per-request RPC-DataGen cost

excelle08
authored and
meta-codesync[bot]
committed
4bf254c
Add mock story-processor module to close prod multifeed Ranking-Story gap

excelle08
authored and
meta-codesync[bot]
committed
6330022
Make prod-aligned knobs the defaults (#712)

excelle08
authored and
meta-codesync[bot]
committed
45709d9