[OpenAFS] client statistics and metrics and general design

Jakub Moscicki jakub.moscicki@cern.ch
Thu, 9 Jun 2011 13:27:11 +0200


--0015175cdaecfa5e3704a545bb1c
Content-Type: text/plain; charset="ISO-8859-1"

Hello,

I've been searching in the archives for this bit of information but without
much success.

I'd like to monitor and analyse the end-to-end time for various AFS
operations (open file, read file, list directory) from the point of view of
the client and compare it against
the statistics we collect on the servers. This is because 1) I believe that
end-to-end client times ultimately correspond to the QoS perceived by users
and 2) this could help to disentangle
server-side timing issues from anything that happens in between the client
and server (including various AFS-internal components, cache performance and
ultimately fabric: network, host load etc). I have an impression that the
fabric monitoring (which we obviously already have) is not enough.

The xstat_cm_test tool reports many metrics but they are sparsely documented
(e.g. http://docs.openafs.org/AdminGuide/apc.html#HDRWQ618). Could someone
shed some light on how to best get the numbers for my use-case?

Another approach which I took was to analyse the sources. As I am new to
AFS, I wanted to have a general idea about the design of the client and
tried to generate call graphs with tools like cflow for the AFS kernel
module, without convincing results yet. Is there any documentation on the
internal of client design (linux) and possibly some diagrams or
tools/scripts/resources to get them? Anything that could be useful to get a
better understanding of how client internally works and which are the
relevant available statistics?

Many thanks.

-- 
---
Best regards,
Kuba

--0015175cdaecfa5e3704a545bb1c
Content-Type: text/html; charset="ISO-8859-1"
Content-Transfer-Encoding: quoted-printable

Hello,<div><br></div><div>I&#39;ve been searching in the archives for this =
bit of information but without much success.</div><div><br></div><div>I&#39=
;d like to monitor and analyse the end-to-end time for various AFS operatio=
ns (open file, read file, list directory) from the point of view of the cli=
ent and compare it against</div>
<div>the statistics we collect on the servers. This is because 1) I believe=
 that end-to-end client times ultimately correspond to the QoS perceived by=
 users and 2) this could help to disentangle</div><div>server-side timing i=
ssues from anything that happens in between the client and server (includin=
g various AFS-internal components, cache performance and ultimately fabric:=
 network, host load etc). I have an impression that the fabric monitoring (=
which we obviously already have) is not enough.</div>
<div><br></div><div>The xstat_cm_test tool reports many metrics but they ar=
e=A0sparsely=A0documented (e.g. <a href=3D"http://docs.openafs.org/AdminGui=
de/apc.html#HDRWQ618">http://docs.openafs.org/AdminGuide/apc.html#HDRWQ618<=
/a>). Could someone shed some light on how to best get the numbers for my u=
se-case?</div>
<div><br></div><div>Another approach which I took was to analyse the source=
s. As I am new to AFS, I wanted to have a general idea about the design of =
the client and tried to generate call graphs with tools like cflow for the =
AFS kernel module, without convincing results yet. Is there any documentati=
on on the internal of client design (linux) and possibly some diagrams or t=
ools/scripts/resources to get them? Anything that could be useful to get a =
better understanding of how client internally works and which are the relev=
ant available statistics?</div>
<div><br></div><div>Many thanks.</div><div><br>-- <br>--- <br>Best regards,=
<br>Kuba<br>
</div>

--0015175cdaecfa5e3704a545bb1c--