On 11 Jan 00, at 18:42, Michael S. Keller wrote:
- Performance I'm getting? For sequential writes, like when I
create an oracle tablespace, I get between 15-20 MB/s. Random writes
NFS v2 and v3 handle writes differently. Which version do you use, do you use UDP or TCP and with what packet sizes?
I don't remember the specifics of all the combinations I tried, but I played with v2, v3, udp and tcp. I'm currently running v3 over udp with (i think) 32k blocksize. Performance didn't vary much reguardless of the settings - this supprised me. This is running across giga-bit ethernet with it's 1500 byte packet size limit.
When My F740 is running at 15-20mb/s it's cpu is bouncing around between 80% to 100%, so I think this is the practical sequential i/o load it can handle.
How many mount points do you use?
I've got about 6 mount points, but I have all the db files under one mount point.
NFS will block if you run everything through one mount point.
I don't understand this, but then there's so much I don't understand about NFS. This is something else Netapp needs to consider when selling a Netapp as a db disk subsystem. Were fairly good at using disk subsystems, but we have little experience with NFS.
Anyway, are you saying that a NFS mount point can only have one outstanding i/o across it at any point? From what I read each biod could have one outstanding i/o although the manual didn't say anything about mount points.
Thanks
Rick
---------------------------------------------------------------------- Richard L. Rhodes e: rhodesr@firstenergycorp.com Ohio Edison Co. p: 330-384-4904 f: 330-384-2514
NFS will block if you run everything through one mount point.
I don't understand this, but then there's so much I don't understand about NFS.
This is a client-side architecture dependent issue, but for many unix systems, nfs requests for each individual mount point are serialised. This can create a per-mount point bottleneck, so the obvious way around the problem is to use multiple mount points, which can parallelize things to a degree.
Nick
On Wed, 12 Jan 2000, Nick Hilliard wrote:
This is a client-side architecture dependent issue, but for many unix systems, nfs requests for each individual mount point are serialised.
I think this could certainly be true of NFS via TCP, unless TCP for transactions is used.
Tom
"Richard L. Rhodes" wrote:
NFS will block if you run everything through one mount point.
I don't understand this, but then there's so much I don't understand about NFS. This is something else Netapp needs to consider when selling a Netapp as a db disk subsystem. Were fairly good at using disk subsystems, but we have little experience with NFS.
Anyway, are you saying that a NFS mount point can only have one outstanding i/o across it at any point? From what I read each biod could have one outstanding i/o although the manual didn't say anything about mount points.
I base my comments on notes from bCandid, who has extensive experience supporting Usenet News running on NFS. bCandid recommends many mounts and my experience supports that position.
Try distributing your data files across two mounts and place them so that I/O should be evenly balanced between the two mount points. See if load at the filer rises. See if client performance improves. I make no promises for your environment.
-
Michael S. Keller -
Nick Hilliard -
Richard L. Rhodes -
tkaczma＠gryf.net