> I have done such tests but I don't have them at hand to share with you > at the moment. Expensive single disks such as 15k SAS can do 1Gb/s. But > your typical 7.2k RPM SATA does 50-70 it seems. I will see if I can dig
You are confusing bits & bytes. 15k SAS may only do 1Gbit/sec, but your typical 7.2k SATA will do 400-500MBit/sec (which is, I agree, 50-70 MByte/sec, provided you acknowledge that's BYTES, not BITS) I just measured some disks sustained write performance. The results were: (All disks are sata 7200rpm) My laptop: 457-468 Mbit/sec sustained Single disk workstations/servers: 395-436 Mbit/sec sustained Mirrored disks: 459 Mbit/sec sustained Raid5 disks: 1.43 Gbit/sec sustained If I had done this across a network, it would only take 2 disks to max out the Ethernet (or 1 disk if it's 15k SAS and if the aforementioned stats are correct for the 15k SAS disks), and if I had infinitely bonded Ethernet jacks, it would only take 6 disks to max out the internal storage bus of the storage manager (assuming 3Gbit bus) Given even a tiny amount of parallel computation nodes that are IO intensive, decentralized local storage alleviates the bottleneck at low cost. > > Yes, all the compute nodes will be working in parallel, that's the > > whole point. They will all be writing to local scratch space, and > > consuming a lot of cpu and ram at the same time. > > That is a rather different use case than mine in which case you > probably have no choice but to shell out very big bucks for > fiberchannel or better. Nope. Just use local disks attached to each compute node. Decentralize all the scratch space. Small bucks, and easy. _______________________________________________ Tech mailing list [email protected] http://lopsa.org/cgi-bin/mailman/listinfo/tech This list provided by the League of Professional System Administrators http://lopsa.org/
