LHC Sixtrack News
Status, Sunday 9th September, 2012.
All running well still. One user reports "Maximum Elapsed Time Exceeded" though on several, all? of his, WUs. Still checking for MacOS results but no further complaints at the moment. I present some basic info. There have been several changes to URLs and Servers outwith my control. The correct site is http:lhcathomeclassic.cern.ch/sixtrack/ This can indeed be found easily from LHC@home and then The Sixtrack Project (rather than Test4Theory). The current server is email@example.com. I define "normal" WUs as 10**5/100,000 turns but remember all particles may be lost after an arbitrary number of turns, sometimes, even just a few turns at large amplitudes. Long WUs are 10**6 or one million turns and very Long WUs 10**7 or 10 million turns, and who knows maybe one day 10**8 turns. That depends on how the floating-point error accumulates and at which point the loss/increase of energy and loss of symplecticity invalidate the results. It will be exciting to find out. For Functionality, Reliability and Performance. While waiting for the LXTRACK user node and the second server for test and backup (I assume they will finally get approved!): Functionality; adequate for the moment. It would be good to have a priority system, three levels. 1. Run first, after other Level 1. 2. Normal; queue after Level 1 and before Level 3. 3. Run only if No Level 1/2 tasks queued. I am thinking in terms of running 10**7 jobs as a series of 10**6 jobs. This requires returning and submitting more data, the fort.6 output and the checkpoint/restart files as a minimum. This would be very good additional functionality in itself. Reliability; pretty good but needs the backup server, LXTRACK, and less reliance on CERN AFS.. Should provide a quick test (1 or 2 minutes) to verify the node produces correct results without running the whole WU. This would not obviate result validation but would avoid wasting resources. I could also provide a longer test on the WWW with canonical results that any volunteer could run if he suspects he has over-clocked or is getting results rejected. Performance; pretty good now with SSE2, SSSE3, PNI or whatever. Should implement GPU option. Should measure the cost of the numeric portability. (Incidentally Intel are hosting a Webinar on this topic on Wednesday, but I guess it will address only Intel H/W.)