Differences

This shows you the differences between two versions of the page.

--- cluster:192 [2020/02/24 19:21]
hmeij07
+++ cluster:192 [2020/02/27 15:46]
hmeij07 [Miscellaneous]
@@ Line 4: / Line 4: @@
 ===== EXX96 =====
-A page for me on how these 12 nodes were build up after they arrived. To make them "ala n37" which as the test node in redoing our K20 nodes, see [[cluster:172|K20 Redo]]
+A page for me on how these 12 nodes were build up after they arrived. To make them "ala n37" which was the test node in redoing our K20 nodes, see [[cluster:172|K20 Redo]] and [[cluster:173|K20 Redo Usage]]
-Page best read bottom to top.
+Page best followed bottom to top if interested in the whole process.
+The Usage section below is HPCC users wnatig to use queue ''exx96''.
+==== Usage ====
+The new queue ''exx96'' will be comprised of nodes ''n79-n90''.  Each node holds 4x RTX2080S gpus, 2x Xeon Silver 4214 2.2 Ghz 12-core cpus, 96 GB memory and a 1TB SSD. ''/localscratch'' is around 800 GB.
+A new static resource is introduced for all nodes holding gpus. ''n78'' in queue ''amber128'' and ''n33-n37'' in queue ''mwgpu'' and the nodes mentioned above.  The name of this resource is ''gpu4''.  Moving forward please use it instead of ''gpu'' or ''gputest''.
+The wrappers provided assume your cpu:gpu ratio is 1:1 hence in your submit code you will have ''#BSUB -n 1'' and in your resource allocation line ''gpu4=1''.  If your ratio is something else you can set CPU_GPU_REQUEST. For example CPU_GPU_REQUEST=4:2 expects the lines ''#BSUB -n 4'' and ''gpu4=2'' in your submit script. Sample script at ''/home/hmeij/k20redo/run.rtx''
+The wrappers (n78.mpich3.wrapper for ''n78'', and n37.openmpi.wrapper for all others) are located in ''/usr/local/bin'' and will set up your environment and start either of these applications: amber, lammps, gromacs, matlab and namd from ''/usr/local''.
+<code>
+# command that shows gpu reservations
+bhosts -l n79
+             gputest gpu4
+ Total             0    3
+ Reserved        0.0  1.0
+# old way of doing that
+lsload -l n79
+HOST_NAME               status  r15s   r1m  r15m   ut    pg    io  ls    it   tmp   swp   mem    gpu
+n79                         ok   0.0   0.0   0.0   0%   0.0     0   0 2e+08  826G   10G   90G    3.0
+</code>
+Peer to peer communication is possible (via PCIe rather than NVlink) with this hardware.  This will get rather messy in setting up.  Some quick off the cuff performance data reveals some impact. Generally in our environment the gains are not worth the effort.  Using Amber and ''pmemd.cuda.MPI''
+<code>
+                                                                              cpu:gpu
+mdout.325288:|  Master Total CPU time:          982.60 seconds     0.27 hours   1:1
+mdout.325289:|  Master Total CPU time:          611.08 seconds     0.17 hours   4:2
+mdout.326208:|  Master Total CPU time:          537.97 seconds     0.15 hours  36:4
+</code>
 ==== Miscellaneous ====
-Add scheduler, reconfigure, elim. Test it out.
+Install scheduler RPM for CentOS7, reconfigure (hosts, queue, static resource), elim. Test it out with old wrapper.
 Edit the n37.openmpi.wrapper for n33-n37 and n79-90 and the one on n78 for the new static resource ''gpu4''.
-Add nodes to ZenOSS hpcmon
+Add nodes to ZenOSS hpcmon.
+Propagate global ''known_hosts'' files in users ~/.ssh/ dirs.
+Look at how accounting ties in with resource request ''gpu4='' versus ''gpu='' ...
 <code>
@@ Line 43: / Line 86: @@
 # for mwgpu/exx96 -pm=ENABLED -c=DEFAULT
-nvidia-smi --persistence-mode=1
+# note: turned this off, running with defaults
-nvidia-smi --compute-mode=0
+#nvidia-smi --persistence-mode=1
+#nvidia-smi --compute-mode=0
 # turn ECC off (memory scrubbing)

DokuWiki

User Tools

Site Tools

Differences

Page Tools