Differences

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

--- cluster:133 [2014/08/08 19:36] – hmeij
+++ cluster:133 [2015/03/18 18:26] (current) – [High Core Count - Low Memory Footprint] hmeij
@@ Line 2: / Line 2: @@
 **[[cluster:0|Back]]**
-==== High Core Count Low Memory Footprint ====
+==== High Core Count - Low Memory Footprint ====
 I polled some folks with the problem described below to find a solution. Then ...
@@ Line 9: / Line 9: @@
 We're on the cusp of a new era!
+Other solutions than the one described below
+  * Amax 4U/288 cores [[http://www.amax.com/hpc/product.asp?value=High%20Density%20/%20Performance]]
+  * Microway 2U/144 cores [[http://www.microway.com/products/hpc-clusters/high-performance-computing-with-intel-xeon-hpc-clusters/]]
+==== Ideas ====
+One idea I received back was to look at the Intel Atom line of chips. From Andrew
+"We can definitely quote rackmounted Atom servers in fairly dense configurations. One example of what we could quote would be : Within each 3U enclosure :12x Sleds, each with TWO C2750 Atom systems on it. So per 3U box :: 24x C2750 Atom systems, each can have 2x 2.5" HDD, Up To 64GB Memory, 2x 10/100/1000 NIC, VGA Port".
+That's a 4-core chip (quoted) so 96 cores/3U. Could double soon with 8 core chip.
+  * Intels calls this design "microservers". From Tower, to rack, to blade, to microservers.
+  * Details at [[http://www.intel.com/content/www/us/en/servers/microservers.html]]
+  * [[http://newsroom.intel.com/community/intel_newsroom/blog/2013/09/04/intel-unveils-new-technologies-for-efficient-cloud-datacenters|Intel Unveils New Technologies for Efficient Cloud Datacenters]]
+So I went looking at my favorite vendor's hardware platform and found:
+{{:cluster:microbade.jpg?200|}}
+[[http://www.supermicro.com/products/MicroBlade/|MicroBlade!]] 896 cores in 6U. Ok then.
+  * 28 blades, 112 nodes, 4 nodes per blade, each node with
+    * 1x Atom C2750 8 core 2.4 Ghz chip
+    * up 32 GB ram (4 GB per core, way above what's needed)
+    * 1x 2.5" disk
+  * Virtual Media Over LAN (Virtual USB Floppy / CD and Drive Redirection)
+  * Do these PXE boot? How to get OS on drives?
+  * Other thoughts
+    * With that many nodes, /home would probably not be mounted
+    * So users would have to stage job data in /localscratch/JOBPID probably
+    * ... via scp from a target host
+==== Slurm ====
+And then we need something that can handle ten of thousand of jobs if we acquire such a dense core platform.
+Enter [[https://computing.llnl.gov/linux/slurm/|Slurm]], which according to their web site, "can sustain a throughput rate of over 120,000 jobs per hour".
+Now we're talking.
+Notes on Slurm are [[cluster:134|High Core Count - Low Memory Footprint]]
 ==== Problem ====