[Beowulf] Q: IB message rate & large core counts (per node)?
Many of your questions may have already been answered in earlier discussions or in the FAQ. The search results page will indicate current discussions as well as past list serves, articles, and papers.
Brian Dobbins bdobbins at gmail.comTue Feb 23 14:35:41 PST 2010
- Previous message: [Beowulf] Q: IB message rate & large core counts (per node)?
- Next message: [Beowulf] Q: IB message rate & large core counts (per node)?
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Hi Patrick, I have been quite vocal in the past against the merit of high packet rate, > but I have learned to appreciate it. There is a set of applications that can > benefit from it, especially at scale. Actually, packet rate is much more > important outside of HPC (where application throughput is what money buys). > The 'especially at scale' bit seems to me to be the critical issue - weighing the price/performance as the ratio of small-scale to large-scale runs changes, assuming that an adapter with better large-scale performance has a significant cost differential. If only we knew what that ratio would be ahead of time, this would be easier. :-) However, I would pay attention to a different problem with many-core > machines. Each user-space process uses a dedicated set of NIC resources, and > this can be a problem with 48 cores per node (it affects all vendors, even > if they swear otherwise). You may want to consider multiple NICs, unless you > know that only a subset of the cores are communicating through the network > (hybrid MPI/Open-MP model for example) or that the multiplexing overhead is > not a big deal for you. Well, clearly we hope to move more towards hybrid methods -all that's old is new again?- but, again, it's *currently* hard to quantify the variables involved. Time to transition, performance differences, user effort, etc. But getting back to a technical vein, is the multiplexing an issue due to atomic locks on mapped memory pages? Or just because each copy reserves its own independent buffers? What are the critical issues? > You need PCIe Gen2 x16 to saturate a 32 Gb/s QDR link. There is no such NIC > on the market AFAIK (only Gen1 x16 or Gen2 x8). But even then, you won't > have any PCIe bandwidth left to drive a second port on the same NIC. There > may be other rationales for a second port, but bandwidth is not one of them. > I thought PCIe Gen2 x 8 @ 500 Mhz gives 8GB/s? I know there are 250 and 500 Mhz variants in addition to the lane sizes, so while a 250 Mhz x8 link wouldn't provide enough bandwidth to a dual-port card, the 500 Mhz one should. But I'm woefully out of date on my hardware knowledge, it seems. Of course, EDR (eight data-rate) IB is on the roadmap for 2011, so if we're in no rush that could help, too. Cheers, - Brian -------------- next part -------------- An HTML attachment was scrubbed... URL: http://www.scyld.com/pipermail/beowulf/attachments/20100223/6911c38f/attachment.html
- Previous message: [Beowulf] Q: IB message rate & large core counts (per node)?
- Next message: [Beowulf] Q: IB message rate & large core counts (per node)?
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
More information about the Beowulf mailing list
