Hi, Shawn. I don't realy follow the thread, but want to comment on > I think most network filesystems (NFS,SMB, etc) cannot be locally cached. I played with EFS recently. It seems to be cached locally pretty much. After, a searcher opens a mounted index directory, reads rate spikes for some time, but then, until the commit, read rate is neglectable.
On Wed, Jul 6, 2022 at 10:59 PM Shawn Heisey <apa...@elyograg.org> wrote: > On 7/6/22 10:59, dmitri maziuk wrote: > > mmap() doesn't side-step disk access though, dep. on the number of of > > mmap'ed chunks and chunk size, it can be slow. Especially if your > > "disk" is an iSCSI volume on a gigabit link to a slow underprovisioned > > NAS. > > If the mounted filesystem is one that the OS can cache, and there is > enough spare memory, then a lot of the mmap requests that Lucene makes > won't ever hit the actual disk. Most block devices can be cached. I > would expect that to be the case for iSCSI, because it should be a block > device from the OS perspective. I think most network filesystems (NFS, > SMB, etc) cannot be locally cached. They are probably cached on the > server side, but then you'd be limited by network bandwidth and > latency. The transfer rate of most 7200RPM SATA disks is a little bit > faster than gigabit ethernet. > > Thanks, > Shawn > > -- Sincerely yours Mikhail Khludnev
