TeraGrid '08, Las Vegas, NV

The annual TeraGrid Users' meeting is in Las Vegas this week. Since SuperComputing was in Reno last November and now the TeraGrid meeting is here, it appears that either Nevada is the hub of the HPC universe, or people who build big computers like to gamble.

There were tutorial sessions yesterday. I attended the Data Management and Visualization sessions. The overall theme seemed to be that the biggest problem associated with using TeraGrid resources is moving large data sets around. The Data Management session in the morning outlined ways to move data to a TG compute resource. "Your mileage may vary" was the big theme. It sounds like you CAN transfer data at 600 MB/sec if you do it just the right way with full knowledge of the file system architecture on both sides, BUT you are more likely to get 1 MB/sec. This sounds like a real opportunity for better software to make this easier and faster. "WAN Lustre" and TeraGrid-aware scp were proposed.

Remote visualization using TG resources was one answer to the data-moving problem. Just don't move it, use it where it is. Kelly Gaither from TACC talked about how to use Maverick (uses Shared Visualization software (VirtualGL, TurboVNC, and Sun Grid Engine) for remote visualization and collaboration and Joe Insley from Argonne National Laboratory talked about using remote ParaView on their graphics cluster. One thing I took away from this session was that parallel visualization applications (like ParaVIew) are really hard for the casual scientific user to use. Also, the mechanics of using the visualization applications remotely seemed complex. At least for TACC, we need to do a better job of educating them on how to use the improvements in Shared Viz 1.1 for ease of use. Paul Navratil talked about using the Ranger cluster for visualization. They have users who use ParaView with software rendering (Mesa). This should be greatly improved in performance when we get the hardware-accelerated graphics cluster attached to Ranger!

This morning's keynote address was by Daniel Reed, currently at Microsoft. It was a though-provoking talk. The main themes were how to deal with the data tsuname, available systems shape research agenda (a corollary of the Saper-Whorf hypothesis that language influences the habitual thought of its speakers), and bulk computing is almost free, but software and power are not, moving data is still hard, people are incredibly expenisve, and robust software remains extremely labor-intensive.
Faster processors enable more new software features which result in slower programs which create demand for even faster processors.
Dr. Reed highlighted the impact that multi-core processors will/are having and the problems this will cause in creating software that can take advantage of them to gain performance.
MPI still dominates, but the level of abstraction needs to be raised. He asserted that purpose of the national investment in computer hardware should be to enable science and not to turn researchers into computer scientists.
The other wave of the future that he focussed on is "cloud computing". Microsoft's "cloud" is growing at an exponential rate. Their new data center in Chicago is just under the 200 MW power envelope. He said the physical plant and power are the real cost, hardware is almost free. Currently, funding agencies pay for hardware acquisition, while the institution pays for physical plant and power. This means that the funding model needs to change.

Comments:

Yes, very true.

Moving data is "the" solution when the compute power is a scares resource. Now the things have changed and people are building Clouds of thousands of computers managed by a single organization. The data can be staged once for the cloud and the analyses will run without moving data in clouds.
Few months back I wrote a paper about this related to the work we did for High Energy Physics Data Analysis. It was not selected by the intended conference, but what we tried to explain was the same concept.
Here is the link if anyone is interested.
http://grids.ucs.indiana.edu/ptliupages/publications/dadc-submission.pdf

Posted by Jaliya Ekanayake on June 23, 2008 at 11:44 PM PDT #

Post a Comment:
  • HTML Syntax: NOT allowed
About

Linda Fellingham

Search

Categories
Archives
« July 2014
SunMonTueWedThuFriSat
  
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
  
       
Today