26 lines
1.2 KiB
Plaintext
26 lines
1.2 KiB
Plaintext
001 isend 送受信に使用するバッファは毎回異なる
|
||
002 barrier
|
||
003 isend 送受信に使用するバッファは一つ、waitの前にsleepしない
|
||
004 isend-calc-wait, all-to-all
|
||
005 lockall-accumulate-calc-unlockall, all-to-all
|
||
006 parent isend-calc-wait, child does nothing --> crash
|
||
007 parent isend-calc-wait, child psm2 send/recv --> one ep per process
|
||
008 parent psm2-init and psm2-connect, child psm2-send/recv --> receiver side crash
|
||
009 parent does nothing, child psm2-init, psm2-connect, psm2-send/recv --> receiver side crash
|
||
010 parent psm2-init, psm2-connect, psm2-send/recv, child does nothing
|
||
011 001にopenmpスレッドを追加
|
||
012 get_acc-calc-flush_local_all, all-to-all. Execute ./012.sh
|
||
013 acc-flush_local-calc, all-to-all, acc:flush_local=1:1
|
||
014 012 + async progress thread.
|
||
015 013 + async progress thread
|
||
|
||
016 MPI_Get_accumulate()のオーバーラップ
|
||
|
||
* 通信パターンは全対全、
|
||
* CPUはいくつかをprogress thread専用に割く
|
||
* ステップは以下の通り
|
||
(1) MPI_Get_accumulate()
|
||
(2) MPI_Get_accumulate()とMPI_Flush_local_all()だけを行った場合の
|
||
時間の0.i倍の計算を実行
|
||
(3) MPI_Flush_local_all()
|