[OpenAFS] Re: [OpenAFS] Re: [OpenAFS] cache manager locked under
Sun, 7 Feb 2010 21:41:35 -0500
2010/2/7 Alena Manova <email@example.com>:
>> ------------ P=F9vodn=ED zpr=E1va ------------
>> Od: Andrew Deason <firstname.lastname@example.org>
>> P=F8edm=ECt: Re: [OpenAFS] Re: [OpenAFS] cache manager locked under heav=
>> Datum: 08.2.2010 02:09:49
>> On Mon, 08 Feb 2010 00:01:03 +0100 (CET)
>> Alena Manova <email@example.com> wrote:
>> > I am running one fileserver process. should I run more? how to get
>> > the optimal number?
>> > Instance fs, (type is fs) currently running normally.
>> > =A0 =A0 Auxiliary status is: file server running.
>> > =A0 =A0 Process last started at Sun Feb =A07 04:00:05 2010 (2 proc sta=
>> > =A0 =A0 Command 1 is '/usr/afs/bin/fileserver'
>> > =A0 =A0 Command 2 is '/usr/afs/bin/volserver'
>> > =A0 =A0 Command 3 is '/usr/afs/bin/salvager'
>> The number of threads you're running with is usually given by the -p
>> option to the fileserver. You are running with the default, which I
>> think is around 9. Try adding the argument "-p 128" (no quotes) to the
>> parameters passed to the fileserver, and see if it affects your problem
>> at all.
> ok, thanks.
> I changed the fileserver to be started with the option -L which means "la=
rge" and implies -p 128.
> however it didn't help solve my cache manager issue.
nor will it.
> I also followed the vCache performance analysis described by Jason but re=
gardless the -stats parameter (even tried with value 1000000) I can't get b=
etter hits/misses ratio than about 50% (which is much worse than recommende=
-stats is not -stat.
> how about the 'dynamic-vcaches' feature? it is enabled by default so is t=
he -stats parameter still relevant?
> I tried to use the -fakestat-all parameter and it seems it can improve th=
e ratio to 4%. but the issue persists (with same scale).
> currently I am running the cache manager with -dynroot -fakestat-all -vol=
umes 256 -daemons 6 -stats 300000 and the fileserver with 128 LWPs. but the=
issue is still there:
as above. with with -stat rather than -stats