charm AT lists.cs.illinois.edu

Subject: Charm++ parallel programming system

List archive

[charm] Question about rank migration efficiency

From: "Van Der Wijngaart, Rob F" <rob.f.van.der.wijngaart AT intel.com>
To: Sam White <white67 AT illinois.edu>
Cc: "charm AT cs.uiuc.edu" <charm AT cs.uiuc.edu>, "Zheng, Gengbin" <gengbin.zheng AT intel.com>, "Georganas, Evangelos" <evangelos.georganas AT intel.com>, "Keppel, Pardo" <pardo.keppel AT intel.com>
Subject: [charm] Question about rank migration efficiency
Date: Thu, 5 Jan 2017 00:38:45 +0000
Accept-language: en-US

Hi Sam,

I used Charm++ v 6.7.1 and built device multicore-linux64-icc for my shared memory machine. I ran my Adaptive MPI code as follows (just an example):

./charmrun [program + arguments] +p36 +vp144 +isomalloc_sync +balancer RefineLB +LBDebug

I tested this with explicit Pack/Unpack, and without (both giving the correct results, of course). On the SMP node the former was 3x slower than the latter. We surmised that the one without PUP (so using only isomalloc) was more efficient because it does not actually move any data, just pointers. Is that correct? Thanks!

Rob

[charm] Question about rank migration efficiency, Van Der Wijngaart, Rob F, 01/04/2017