[OpenAFS] performance and udp buffers

Jakub Moscicki Jakub.Moscicki@cern.ch
Tue, 20 Nov 2012 08:29:44 +0000


Hi Simon,

> On 19 Nov 2012, at 07:15, Jakub Moscicki wrote:
>> Thanks for this analysis. The increased UDP works pretty well for us at =
CERN so far - albeit one limit gone other limits appear more pronounced.
>=20
> I'm interested in what other limits you are hitting. I'm very aware of th=
e problems with the listener thread load and scheduling, are you hitting an=
y other problems with RX, or are they fileserver limitations?

Actually both. It's hard to fill up the 10GE network with RX for some strea=
ming use-cases from multiple underlying drives (not a major limitation for =
us yet but might be useful for e.g. volserver volume moves). Conversly, wit=
h UDP packet loss giving better performance, it is much easier to saturate =
an underlying individual drive in case of many clients of a single user ham=
mering one volume (batch jobs essentially). Two possible paths: more effici=
ent caching (with the SSD layer via device mapper or with just more RAM for=
 buffers) and smarter throttling in the fileserver (for example, scheduling=
 of worker threads to take into account underlying I/O limitations and/or a=
ccording to QoS to be provided by the fileserver [e.g. home directory files=
erver geared towards interactive use versus workspace fileserver geared tow=
ards batch jobs]).  Both have many open questions. Will keep you posted.


>=20
>>> management packets. 16Mbytes should be plenty providing that you don't
>>>=20
>>> d) Have a large number of 1.6.0 clients on your network
>>=20
>> Do you mean 1.6.0 (referring to a specific bug in 1.6.0) or 1.6.x (refer=
ring to some general change in client behaviour in 1.6 series)?=20
>=20
> Specificaly 1.6.0, and prereleases. There is a truly unfortunate bug in t=
hose clients which causes them to create a gradually escalating ping flood =
against every fileserver they contact. At its worst, this creates a distrib=
uted denial of service attack against your fileservers. In terms of this di=
scussion, the large number of incoming RX version packets can overwhelm the=
 listener thread. As these packets are not flow controlled, they can force =
"real" data packets out of the UDP buffer. One solution to this problem is =
to drop these packets at the kernel firewall - you want to drop all version=
 packets with an RX epoch of 999.

Good to know! Can this be done with just standard iptables?

kuba

--