Louwtjie Burger writes: > Thanks Andrew, point taken ... but let's say your products ;) are not on the > table as a solution. > > It's a decission between FCIP or IP ... that's it.
Cool. You can still buy our stuff, you'll just want to run it in 10GbE mode ;) > The HPC applications will be a collection of programs, let's think of it as > a "walk-in" HPC centre (ala Sun Grid ...). > > Would the FC not give one the low-latency that most applications desire ... > BUT come with a higher price in getting IP encapsulated into FC (compared to > TCP) ? What is FC/IP's latency on Solaris? Make sure you get the real, through the kernel, userspace to userspace latency, rather than the latency on the wire. The costs of copying in/out of the kernel, scheduling interrupt handlers, and context switching can dwarf the hardware latency in many cases. As an example, our Myri10G nics are "dual personality" and can run in both 10GbE and Myrinet modes. The same hardware which gives you a 2.34us 1/2 rtt latency in OS-bypass (Myrinet) mode gives you a much higher latency when you go through the kernel. On linux on opteron, our 10GbE latency is something like 9us, on Solaris on opteron, it is on the order of 14us. Drew _______________________________________________ networking-discuss mailing list [email protected]
