[Beowulf] Cheap SDR IB
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.
Gilad Shainer Shainer at mellanox.comThu Jan 31 11:32:56 PST 2008
- Previous message: [Beowulf] Cheap SDR IB
- Next message: [Beowulf] Cheap SDR IB
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Richard Walsh wrote: >> With more cores on a single node, the IB benefits are seen in much lower number >> of nodes. I am testing some applications on a new cluster that I have (dual >> sockets quad core Barcelona), and my first results are with Fluent new >> benchmarks. I will have the numbers posted soon, so you all can take a look. For >> 2 nodes, IB shows an average of 15-20% higher performance then GigE, and this >> gap gets bigger with cluster size. At 4 nodes the difference was 40-50%. Even >> more important, 3 nodes results with IB were higher then 8 nodes with GigE, and >> GigE stop scaling after 3-4 nodes (performance numbers were flat after 3-4 >> nodes). > Interesting. I am guessing this is with ConnectX adapters and PCIe Gen2. When does > the InfiniBand technology used in this test flatten? How about the less expensive SDR > technology referred to in Jeff's article? Can you provide the curves up to 8 nodes for > both IB and GE in this test? > > A quick analysis, limited to this Fluent test, suggests that 4 nodes plus ConnectX > technology roughly equals 8 nodes plus on-board GE technology. Can you provide > the system per node price differences? Then we can rougly determine the cost > benefit relationship. In my testing I have used ConnectX and PCIe Gen1. I will have the numbers posted soon and will send a note to this list once done. The major take was that GigE did not scale beyond 3 servers, so adding more nodes did not provide more performance (Fluent rating stayed the same from 3 nodes with GigE and above). The IB shows almost linear scaling as we have added more nodes. The less expensive SDR technology should follow the same lines, as latency was more important in those benchmarks, so the numbers should be in the same range. For 2 nodes only the performance gap was 15-20%. If I am taking the safe approach and assuming 10-15% gap with the low cost SDR. The servers will cost you ~$8K, the IB part (for 2 nodes you don't need a switch) according to the pricing in Jeffs is <$300 (2 adapters and one cable) which is less than 4% added cost. Of course when you scale beyond 2 nodes, the cost/performance advantage with IB increases dramatically (as GigE performance is flat beyond 3 servers).
- Previous message: [Beowulf] Cheap SDR IB
- Next message: [Beowulf] Cheap SDR IB
- Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
More information about the Beowulf mailing list
