Open x41lakazam opened 2 months ago
Mpi test results:
$ mpirun \
-x UCC_CL_HIER_ALLGATHERV_TUNE=allgatherv:@node_split:inf \
-x UCC_CLS=basic,hier \
-x UCC_CL_HIER_NODE_LEADERS_SBGP_TLS=ucp \
-x UCC_CL_HIER_NODE_SBGP_TLS=ucp \
-x LD_PRELOAD=./lib/libucc.so \
-x PATH \
-x LD_LIBRARY_PATH \
-np 6 --map-by ppr:3:node \
./bin/ucc_test_mpi -c allgatherv -t world -d uint32 -v -m 8 -v
===== UCC MPI TEST INFO =======
seed: 6222
collectives: Allgatherv
data types: uint32
memory types: Host
teams: world
tc=Allgatherv team=world mtype=host msgsize=8 persistent=0 inplace=0 dt=uint32
===== UCC MPI TEST REPORT =====
collective tests passed failed skipped
Allgatherv 1 1 0 0
===== UCC MPI TEST SUMMARY =====
total tests: 1
passed: 1
skipped: 0
failed: 0
elapsed: 9s
@x41lakazam is this still draft?
allgatherv algorithm split into three parts:
Still under development. Currently supports only int32, I will add support for int64 once I'm sure of the logic.
Situation where the ranks are not in a continuous order across nodes is not handled yet.