Hi John--
On Sun, 2004-05-30 at 05:31, John Spitz wrote:>
> Lustre throughputs look attractive but I am trying to
> find out more about Lustre''s caching scheme as it will
> directly impact our farm performance:
>
> (a) Do lustre clients cache read content on disk or in
> RAM or both?
Today Lustre clients cache read content only in RAM. We are building
persistent caches and proxies that can also cache on disk.
> (b) If in RAM, can lustre clients on linux 2.6 amd64
> make use of, say most of the 32 GigaBytes of RAM on
> our servers.
Yes. We use the normal Linux page cache on clients, so it will discard
cached data as memory is needed, according to its own policies.
> (c) I saw a document that indicates lustre employs a
> distributed cache where misses are redirected to other
> nodes. Has this been implemented? If so, where can I
> find more information on this - for example, does the
> node where the miss occurred go on to replicate the
> missed content etc?
This has been designed, but not fully implemented -- these are the
proxies that I was referring to. In our early prototype, the proxy node
did indeed replicate the missed content. I expect that the final
version will contain two modes: one in which caches are kept warm at all
times (mirrors), and one in which caches are refreshed on-demand.
I hope that helps. Whether or not you choose to use Lustre, I hope that
you will share your reasons, so that we can continue to improve our
product.
Thanks--
-Phil