[OpenAFS-devel] performance issues with 1.4.8 and above

[Derrick Brashear]

We should probably scream this out a bit more; the unfortunate thing
is because of what it does to guess MTU, it's really easy to end up
with fragmented packets, and if your network configuration results in
fragments being dropped (and with clients at the edge, it's not likely
to be strongly under most cells' control) things get ugly fast.


On Thu, Apr 9, 2009 at 8:42 AM, Felix Frank <Felix.Frank@desy.de> wrote:
> On Thu, 9 Apr 2009, Derrick Brashear wrote:
>
>> On Thu, Apr 9, 2009 at 8:19 AM, Felix Frank <Felix.Frank@desy.de> wrote:
>>>
>>> Hi,
>>>
>>> I've been busy doing a couple of stress tests lately. A disturbing
>>> finding
>>> is that the 1.4.7 fileserver achieves nearly 25% higher data throughput
>>> than
>>> 1.4.8 and 1.4.10 respectively.
>>
>> My suspicion: force jumbograms to be on for your environment.
>>
>> The issue was they could not previously be turned off, regardless of
>> switch, due to a programming error in OpenAFS since 1.0.
>
> Good call. The throughputs are identical after including -jumbo for 1.4.10.

Derrick