From: Douglas Houston (DouglasR.Houston_at_ed.ac.uk)
Date: Thu Sep 18 2014 - 05:35:30 CDT
We have a 802.3ab 1000BASE-T Gigabit Ethernet switch (Netgear GS205)
that is dedicated to connecting the nodes of this cluster only (so no
additional traffic). Assuming removing this switch will allow two
nodes to show a speedup relative to one node, would this really
represent a solution?
It was my understanding that NAMD is at least somewhat parallelisable
across an ordinary ethernet-linked cluster of nodes. Is this not the
Can anyone on this mailing list tell me if they have successfully
noted a speed increase by running NAMD across an ethernet-linked
cluster vs. a single node? If so, could they please list their network
hardware, e.g. ethernet cards, switch, number of nodes, cores per
node, size of system benchmarked, etc.
P.S. 64 minutes of simulation runtime across 2 of the nodes results in
a total of 390GiB of data transferred between them (according to
ifconfig) - this equates to about 100 MiB/sec. This is for a
80,000-atom system. For my small 5,000 atom system it shows about 60
MiB/sec. Does this mean that, for the large system at least, the
bandwidth could indeed be saturating (100 MiB/sec being not far off
1Gbit/sec)? If this is the case, it is not clear to me why the data
transfer rate is so much and if anything can actually be done about it.
Quoting Norman Geist <norman.geist_at_uni-greifswald.de> on Wed, 10 Sep
2014 15:14:37 +0200:
> Benchmark the timing over two nodes, without the switch. Sometimes the
> switches are very slow, especially if other ports are active at the same
> Norman Geist.
Dr. Douglas R. Houston
Institute of Structural and Molecular Biology
Room 3.23, Michael Swann Building
University of Edinburgh
Edinburgh, EH9 3JR, UK
Tel. 0131 650 7358
-- The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.
This archive was generated by hypermail 2.1.6 : Wed Dec 31 2014 - 23:22:51 CST