Lustre discuss - May 2006 - Sizing up memory

On Tue, 2004-06-29 at 02:59, João Miguel Neves wrote:
> Is there a rule of thumb for sizing up memory for the MDS on lustre
> 1.0.4?
> 
> I''m always getting messages that indicate lack of memory:
> 
> __alloc_pages: 0-order allocation failed (gfp=0x20/0)
> 
> So my plan is to get more memory, but how much?
> 
> If there''s something I can tune to reduce the memory usage,
I''d be very
> thankful.
There is no fundamental problem with using an MDS with a small amount of
ram.  Lustre uses a few 100k per client connected, and about 1MB per
1,000 locks.  Any extra ram will be used by the kernel to cache inodes
and directories and so on.  What you are reporting sounds like a problem
that we have encountered in the past -- although never on the MDS --
with keeping enough "emergency memory" free to handle unexpected
requests.

Are you running only an MDS on this node, or is it running other
applications or a Lustre client?

It is almost certainly strongly related to the kernel that you are
running.  I seem to recall that the kernel.org kernels have many more
problems in this regard than the Red Hat kernels, but I could be
mistaken.  A few more details about your configuration will be helpful.

Thanks--

-Phil

--=-ixk9T7h/K8RLjysNU3vb
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: quoted-printable

Sorry. Yes, the machine is both an MDS and a client.

Yes, this is a vanilla kernel with the patches applied with quilt.

The messages seem to be related with portals (at least that''s where I
get most of the messages).

Any suggestions on how to reduce these effects or any information you
have would be great. Thanks.

A Ter, 2004-07-13 =C3=A0s 04:43, Phil Schwan escreveu:
> On Tue, 2004-06-29 at 02:59, Jo=C3=A3o Miguel Neves wrote:
> > Is there a rule of thumb for sizing up memory for the MDS on lustre
> > 1.0.4?
> >=20
> > I''m always getting messages that indicate lack of memory:
> >=20
> > __alloc_pages: 0-order allocation failed (gfp=3D0x20/0)
> >=20
> > So my plan is to get more memory, but how much?
> >=20
> > If there''s something I can tune to reduce the memory usage,
I''d be very
> > thankful.
>=20
> There is no fundamental problem with using an MDS with a small amount of
> ram.  Lustre uses a few 100k per client connected, and about 1MB per
> 1,000 locks.  Any extra ram will be used by the kernel to cache inodes
> and directories and so on.  What you are reporting sounds like a problem
> that we have encountered in the past -- although never on the MDS --
> with keeping enough "emergency memory" free to handle unexpected
> requests.
>=20
> Are you running only an MDS on this node, or is it running other
> applications or a Lustre client?
>=20
> It is almost certainly strongly related to the kernel that you are
> running.  I seem to recall that the kernel.org kernels have many more
> problems in this regard than the Red Hat kernels, but I could be
> mistaken.  A few more details about your configuration will be helpful.
>=20
> Thanks--
>=20
> -Phil
>=20
> _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss@lists.clusterfs.com
> https://lists.clusterfs.com/mailman/listinfo/lustre-discuss
--=20
						Jo=C3=A3o Miguel Neves



--=-ixk9T7h/K8RLjysNU3vb
Content-Type: application/pgp-signature; name=signature.asc
Content-Description: Esta =?ISO-8859-1?Q?=E9?= uma parte de mensagem
	assinada digitalmente

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (GNU/Linux)

iD8DBQBA9BVMGFkMfesLN9wRApWfAJwIyGk3OccG4LiXzscGXIJ1bu15wQCfTnny
oKr3nP+z5KGnIJ+rjmW6W+E=ZNnJ
-----END PGP SIGNATURE-----

--=-ixk9T7h/K8RLjysNU3vb--

--=-OOBE419ejXpcKeLHJTib
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: quoted-printable

Is there a rule of thumb for sizing up memory for the MDS on lustre
1.0.4?

I''m always getting messages that indicate lack of memory:

__alloc_pages: 0-order allocation failed (gfp=3D0x20/0)

So my plan is to get more memory, but how much?

If there''s something I can tune to reduce the memory usage,
I''d be very
thankful.
--=20
						Jo=C3=A3o Miguel Neves



--=-OOBE419ejXpcKeLHJTib
Content-Type: application/pgp-signature; name=signature.asc
Content-Description: Esta =?ISO-8859-1?Q?=E9?= uma parte de mensagem
	assinada digitalmente

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (GNU/Linux)

iD8DBQBA4RNqGFkMfesLN9wRApxuAJ43z4vHAI56xkEIID+0CBzKeQGCOQCgheyN
OTaYhrb0fxqZKPaaPS+5q5g=/UiX
-----END PGP SIGNATURE-----

--=-OOBE419ejXpcKeLHJTib--