Optimizations to multi-process and multi-rank Datadeps - #739

Merged

jpsamaroo merged 26 commits into

masterfrom

jps/hierarchical-mp-mr

Sep 17, 2026

Commits on Sep 17, 2026

datadeps/hierarchical: Submit off the planning thread, and measure planning

jpsamaroo
and
cursoragent
committed
02dfd4b
Datadeps: Memoize, batch and defer aliasing round-trips

jpsamaroo
and
cursoragent
committed
84786b3
Benchmark CI: Add per-job regression/improvement summary, bump default samples

jpsamaroo
and
claude
committed
3f846fb
MPIExt: Judge cross-rank waits by progress, not by elapsed time

jpsamaroo
and
claude
committed
410695b
datadeps/hierarchical: Replace HIER_TIMING with TimespanLogging categories

jpsamaroo
and
claude
committed
e0f813a
benchmark/stencil: Await every asynchronous fixture
jpsamaroo
committed
ffbbab6
benchmark/mpi: Synchronize SPMD sampling across ranks
jpsamaroo
committed
223eb88
benchmark/mpi: Support BenchmarkTools 1.8 sampling
jpsamaroo
committed
a6d065f
datadeps: Keep synchronous submission batches short
jpsamaroo
committed
8ea4c90
benchmark: Await vector fixtures before sampling
jpsamaroo
committed
eb8d98b
datadeps: Join disjoint remainder copies before whole use
jpsamaroo
committed
4005db6
benchmark: Report the MPI leaf that aborts
jpsamaroo
committed
ec23266
benchmark: Stabilize Distributed fixture placement
jpsamaroo
committed
8563030
datadeps: Track every producer of a disjoint copy batch
jpsamaroo
committed
c9f7430
datadeps: Skip clock reads when hierarchical timing is disabled
jpsamaroo
committed
4ad6e59
datadeps: Preserve concrete types for reusable aliasing scratch
jpsamaroo
committed
e2b9d3b
mpi: Serialize each integer uniformity payload only once
jpsamaroo
committed
a3b80f0
reuse: Assert scratch container types inside the macros
jpsamaroo
committed
1977cd5
Reuse remote aliasing batches throughout Distributed planning
jpsamaroo
committed
ec27fc4
Show benchmark change lists immediately after the count summary
jpsamaroo
committed
b37ad6d
Compare Distributed linalg with consistently placed input fixtures
jpsamaroo
committed
2be0122
mpi: Wake only consumers of the delivered broadcast tag
jpsamaroo
committed
8bddfe7
mpi: Allocate receive wait events only for contended streams
jpsamaroo
committed
e92d6c1
reuse: Create pooled task slots only when they are needed
jpsamaroo
committed
974f201
Avoid boxed signature iteration and unused default specialization
jpsamaroo
committed
34a01f7
Scan LFU frequencies without allocating an eviction result
jpsamaroo
committed
6421ac5

Navigation Menu

Uh oh!

Optimizations to multi-process and multi-rank Datadeps - #739

Optimizations to multi-process and multi-rank Datadeps#739
jpsamaroo merged 26 commits into
masterfrom
jps/hierarchical-mp-mr

Commits on Sep 17, 2026

datadeps/hierarchical: Submit off the planning thread, and measure planning

Datadeps: Memoize, batch and defer aliasing round-trips

Benchmark CI: Add per-job regression/improvement summary, bump default samples

MPIExt: Judge cross-rank waits by progress, not by elapsed time

datadeps/hierarchical: Replace HIER_TIMING with TimespanLogging categories

benchmark/stencil: Await every asynchronous fixture

benchmark/mpi: Synchronize SPMD sampling across ranks

benchmark/mpi: Support BenchmarkTools 1.8 sampling

datadeps: Keep synchronous submission batches short

benchmark: Await vector fixtures before sampling

datadeps: Join disjoint remainder copies before whole use

benchmark: Report the MPI leaf that aborts

benchmark: Stabilize Distributed fixture placement

datadeps: Track every producer of a disjoint copy batch

datadeps: Skip clock reads when hierarchical timing is disabled

datadeps: Preserve concrete types for reusable aliasing scratch

mpi: Serialize each integer uniformity payload only once

reuse: Assert scratch container types inside the macros

Reuse remote aliasing batches throughout Distributed planning

Show benchmark change lists immediately after the count summary

Compare Distributed linalg with consistently placed input fixtures

mpi: Wake only consumers of the delivered broadcast tag

mpi: Allocate receive wait events only for contended streams

reuse: Create pooled task slots only when they are needed

Avoid boxed signature iteration and unused default specialization

Scan LFU frequencies without allocating an eviction result