[OpenAFS] Reporting on some recent benchmark results

Simon Wilkinson sxw@inf.ed.ac.uk
Wed, 6 Apr 2011 16:06:42 +0100

On 4 Apr 2011, at 22:18, Garrett Wollman wrote:

> Over the past few days I have performed several benchmarks comparing
> the performance of various OpenAFS server and client configurations.

Thanks for this - it makes for really interesting reading.

The statistic I'm really interested in at present, unfortunately, isn't =
one that you cover. With the imminent release of 1.6.0, what would be =
really interesting to know is a direct comparison between 1.4.14 and =
1.6.0 on the same hardware, for the same workload. I know of workloads =
in which I can clearly show that 1.6.0 is faster, what would be really =
useful is to see, and to understand, is workloads for which it is =

You also mention the performance penalties of fcrypt. One of the =
interesting anomalies of both DES and fcrypt is that as well as being =
weaker than AES, they are also slower. RX using rxgk is significantly =
faster than rx with rxkad. AES is also far more amenable to =
acceleration, allowing the crypto overhead to be transferred to external =

We obviously still have a lot of performance improvements to investigate =
in both the fileserver and in RX. Lock contention in RX is a significant =
problem, over and above the condition variable inefficiencies. I did =
some work on this last autumn for YFS, some of which is now in 1.6.x, =
some of which is only in master, and some of which is still to be =