🚀 go-pugleaf

In article <1993Jan22.221447.10697@odin.corp.sgi.com>  miguel@boytoy.csd.sgi.com  (Michael/Miguel Sanchez) writes:
> In article <JP107.93Jan22121147@grus.cus.cam.ac.uk>, jp107@cus.cam.ac.uk (Jon Peatfield) writes:
> |> Ok, I've had a few nice chats with some of the readers of this group
> |> and with our local SGI support line, who are currently off chasing the
> |> fact that this has been reported to them before but they couldn't find
> |> what had happened last time (what kind of help-desk software does the
> |> support line use?)
> |> 
> |> Anyway it turns out that my SGI Indigos can only push files over NFS
> |> at about 10 to 30 Kbytes/sec (after the patch mentioned below) while
> |> all the rest of our machines (Sun/HP/Mac), can manage over
> |> 110Kbytes/sec (I have a large number of figures if you want them.)  I
> |> got a set of patches from SGI for general network performance (a new
> |> ethernet device driver), which improved performance a good deal, but
> |> still didn't put it in the acceptable range.  After hitting my head
> |> for a while and replacing transceivers and moving machines etc, I
> |> watched the network with etherfind (great tool -- why don't SGI
> |> provide it?) and spotted that the SGIs are waiting after each NFS
> |> write for an acknowledgement from the server.  Our servers don't
> |> acknowledge 'til the block is on physical disk (like normal machines),
> |> so there is a huge wait.  This is as if the biod daemons are not
> |> working (but we have four (4) running.)  On all the other machines in
> |> the dept, if I have N biods then I don't see blocking 'til N+1
> |> requests are outstanding.  The delays get worse when one is over a
> |> high latency netwrok while other machines just cope.  What is worse is
> |> that if the Acknowledgement reply is lost (as sometimes happens on
> |> ethernets (it is UDP after all)), the SGI waits for 4.5 SECONDS before
> |> retrying the send.  A few of those and the time really adds up.
> 
> it almost sounds like you are hard mounting the nfs mount points...are you
> doing that?
> 

Umm, have you tried reading the manual?  (I don't want to be rude, but...)

It isn't clear what part you are replying to here, sorry about the
bandwidth guys!

The only difference that hard/soft mounting makes is that with soft
mounting after RETRANS retransmissions a soft mount will give up and
return an error, while a hard mount will repeat the call forever.

To quote the SGI manual (as that would disagree with any other NFS manual):

    Once the filesystem is mounted, each NFS request waits timeo=n tenths of
    a second for a response.  If no response arrives, the time-out is multi-
    plied by 2 and the request is retransmitted.  When retrans=n retransmis-
    sions have been sent with no reply a soft mounted filesystem returns an
    error on the request and a hard mounted filesystem retries the request.
    Filesystems that are mounted rw (read-write) should use the hard option.
    The number of bytes in a read or write request can be set with the rsize
    and wsize options.

The default timeno is 7, so it should only wait 0.7 sec before
retrying!  I will admit that the 4.5 sec delays were not measured by
myself but by another person a mile down the road who is also unhappy
about the NFS performance so I can't say with hand on heat tht I know
what his setup was... perhaps I should start a club ;-)

In fact all the mounts I've been doing are soft mounts if that did
make any difference!

-- Jon
--
Jon Peatfield, Computer Officer, the DAMTP, University of Cambridge
Telephone: (+44 223) 3-37852     Mail: J.S.Peatfield@amtp.cam.ac.uk

To be consistent with DNS domain ordering, shouldn't news groups have
names like "admin.sun.sys.comp" not "comp.sys.sun.admin" ?  EMWTK