Date: Sat, 6 Oct 2018 08:42:10 -0700 From: Mel Pilgrim <list_freebsd@bluerosetech.com> To: Garrett Wollman <wollman@csail.mit.edu>, freebsd-fs@freebsd.org Subject: Re: ZFS/NVMe layout puzzle Message-ID: <775752d2-1e66-7db9-5a4f-7cd775e366a6@bluerosetech.com> In-Reply-To: <23478.24397.495369.226706@khavrinen.csail.mit.edu> References: <23478.24397.495369.226706@khavrinen.csail.mit.edu>
next in thread | previous in thread | raw e-mail | index | archive | help
On 2018-10-04 11:43, Garrett Wollman wrote: > Say you're using an all-NVMe zpool with PCIe switches to multiplex > drives (e.g., 12 4-lane NVMe drives on one side, 1 PCIe x8 slot on the > other). Does it make more sense to spread each vdev across switches > (and thus CPU sockets) or to have all of the drives in a vdev on the > same switch? I have no intuition about this at all, and it may not > even matter. (You can be sure I'll be doing some benchmarking.) > > I'm assuming the ZFS code doesn't have any sort of CPU affinity that > would allow it to take account of the PCIe topology even if that > information were made available to it. In this scenario, the PCIe switch takes the role of an HBA in terms of fault vulnerability.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?775752d2-1e66-7db9-5a4f-7cd775e366a6>