From: Stepan K. <tw...@me...> - 2014-04-22 09:58:24
|
Thanks, we'll try SCST with iSCSI then, to see how it feels first. It's just that the idea of patching the kernel from svn trunk is a little worrying. Not that it's really different from other kernel option(s), but still. And yes, we're keeping SMP affinity and IO in general in mind, it requires tuning in any case, be it RDMA or iSCSI or FC. On Tuesday, April 22, 2014 11:10:04 Bart Van Assche wrote: > On 04/21/14 13:20, Stepan Kucherenko wrote: > > We're planning to use scst in production environment, first as a standard iSCSI > > target for initial testing and smaller tasks (Intel 1Gb i350 and Mellanox > > Connect-X 3 EN NIC), then later with SRP (over RoCE, Mellanox NICs across > > Mellanox eth switch, so DCB(X) and all that). > > > > How stable scst svn trunk version is for our task ? We can't use 2.6.x or > > 3.1/3.2 Linux kernel so earlier releases are out of the question. > > > > Any specific advice ? > > I think the current SCST svn trunk is a good starting point. SRP over > RoCE support is present on the trunk but not on the SCST 2.2.x branch. > > Please note that achieving optimal performance requires manual tuning > based on a good understanding of the initiator side and target side I/O > stacks. E.g. the I/O scheduler, queue depth and readahead size > configured at the initiator side and also IRQ and thread pinning at the > target side have a significant impact on I/O throughput and latency. > Which parameters are optimal depends on the storage type that is present > at the target side (RAID hard disk array or SSDs). > > Bart. > |