[PVFS2-developers] Metadata performance (II)

[Murali Vilayannur]

Hi Julian,

>
> Meta--CREATE-5Clients-XMeta.pdf and Meta--CREATE-15Clients-XMeta.pdf:
> These diagrams show that the create-rate of the DBPF implementation and 5
> metadata servers increase a lot.
> I'm not sure why performance drops which increasing number of created files
> with tmpfs.
> It may be influenced by the implementation of tmpfs or the behavior of
> berkeley database ? If anyone has a glue why this happens please contact me.

Hmm.. Could it be lock contention on the address space spinlock? But that
would require that multiple threads be pounding to the same file on tmpfs
at the same time....
tmpfs also does swap to a backing device if it overflows RAM, Could it be
because you are hitting the swap devices?


> I think about 2700 creates per second with TAS are quite a lot :-)
> It's a pity that IO limits throughput that much for the DBPF implementation.

Yup. Agreed. Could you elaborate on what TAS does that
dbpf-on-tmpfs-no-sync does not do? If dbpf cannot do well on tmpfs (that
too with no sync!), something ought to be wrong with berkeley db on tmpfs?

Does TAS stick to the same consistency/reliability guarantees that
dbpf-on-tmpfs-no-sync would offer?
Also, for the dbpf experiments, what was the underlying file-system?
If it was ext3, what were the mount options used for journalling? In the
past, we have had big problems with ordered and data journalling modes
(writeback should be okay). If it was some other file system, we would
love to hear which one it was..
Thanks!
Murali