SysMan Mall

Super Computing '06 - Day 4 (yesterday)

Returned back home yesterday leaving SC'06 after a long late night flight from Tampa into San Jose. My target was to go understand how cluster management is being perceived and addressed across the floor by various vendors. A hidden answer appears to be rocks, ganglia and a combination of both along with some home grown brew. Most of the home grown brew suffered from one common malady which is they do best what they understood the most, rest is all on the "road map" :-) and the focus appears to be narrow today. Good news is that we at Sun have most of the heterogenity, extensibility and pluggability worked out and are focussing on the math of large numbers or simply put the scale for the next generation monsterous infrastructure that one will see in the HPC space.

There are many booths in the exhibit section and it was impossible to stop at every stall and do justice so I skipped the ones that did not have the word "cluster management" or related art. I was hoping to see a demo of Insight Manager at the HP booth but what they have is primarily the Microsoft Cluster brew, incidentally the software at display is MS Cluster 2003 (even at the MS booth), funny given it is 2006 Super Computing conference, a lot of dynamics have changed since 2003 :-).  There were other grid management software from folks who make some of the cluster hardware and the most complete I saw was one from Evergrid some things are missing but overall nice. In the list of missing things I felt the capability to add things at bottom (extend) and expose things at top (integrate) were obvious. The most interesting part in their solution was the promise of resource management which I really didn't get much insight into, the gentleman on the floor stated that it does work. Its interesting because the change is not in the kernel and it is difficult to guarantee given scheduling is done at lower levels inside the kernel. The resource management talked about is by the way on Linux only and the interplay if ported to Solaris will be interesting and challenging for them, given its burnt into the Solaris kernel. IBM was not demonstrating any cluster management software or atleast that is the response I got from the person manning the stalls, again I was hoping to see a demo of IBM director or equivalent but it was missing on the booth.

There were other booths where cluster management was being talked about like Cluster Resources,  Verari etc., my conclusion is that rocks and ganglia are the solution which most of the customers will use till an industrial strength management application starts supporting the demands of an HPC environment. My take is that there should be some buzz in the coming years in this area and it will be interesting to see how all those solutions will interplay with each other.

Other notable booth that I found interesting was INRIA France's MPI equivalent for Java. I was obviously probing to get data on cluster management and it was mostly homegrown with certain flavor of Linux in mind. I am back to my day job focussing on System Manager and SunMC, have taken a lot of notes on how to improve our existing products to solve some of the tough problems that will be created as we create bigger HPC clusters.

Posted at 04:40PM Nov 15, 2006 by adikhit in General  |  Comments[0]