[PVFS2-developers]

[·¶Ö¾»ª]

> Hi Zhihua
> Your messages keep getting flagged as spam by our
> mail servers,
> probably because of the yahoo.com.cn domain.  Is it
> possible for you
> to send from another account?  If not, I'll just
> keep an eye on the
> spam folder...

This is from my new mailbox. :)
 
> Neill and Phil tried out the one-file approach and
> really didn't see
> any big performance differences -- even though they
> were pretty sure
> they would.   I'll let them fill in the details, but
> I think you
> should be fine with one database for every file.  
> 
> Do you notice any real improvement with this change?
>  If so, with what
> workloads?
> 
> ==rob

Following tests are all with nosync mode and on one machine.

a new pvfs2 file system
./op_test -d pvfs2:/mnt/pvfs2 -n 1000 -O
22.071590 seconds to open 1000 files: 0.022072 secs/open

a new pvfs2 file system
./op_test -d pvfs2:/mnt/pvfs2 -n 5000 -O
112.024601 seconds to open 5000 files: 0.022405 secs/open

then I change the dbpf.h.

a new pvfs2 file system
./op_test -d pvfs2:/mnt/pvfs2 -n 1000 -O
10.932990 seconds to open 1000 files: 0.010933 secs/open

a new pvfs2 file system
./op_test -d pvfs2:/mnt/pvfs2 -n 5000 -O
52.928119 seconds to open 5000 files: 0.010586 secs/open

So why did you say no improvement?

Berkerly BD manual says that "Calling DB->open is a
relatively expensive operation, and maintaining a set
of open databases will normally be preferable to
repeatedly opening and closing the database for each
new query."
I find that when I call DB->open with
different filenames in my system, the performance is as follows:
1000 files, time: 10.619 seconds(not including the close time)
5000 files, time: 53.853 seconds(not including the close time)

So the main overhead is on DB->open when op_test. Isn't it?