New to NBU ...

Our backups will be of many TB of large (>2GB) files on
ufs file system on Solaris 10, SF25K, well equipped.

The media server is the backup server, using 6.5.4. Master
server is S10 T5220, lots of disk, network traffic between the
two is minimal.

Backup of ordinary stuff is great. 50MB/s on a busy system
through the day, and after caching up the file system with a
few practice backups, it goes 100MB/s!

But - change the input to the sparse files, and it does not do
so well - 10-15MB/s.

The NUMBER_DATA_BUFFERS is 128, the SIZE_DATA_BUFFERS
is set to 256K (262144).

Watching the process w/truss, the bpbkar uses read, with a
read size of 64k of the WHOLE file, one thread, closes the
file it is backing up, then re-opens it, uses pread with up to
20 threads, and feeds that to bptm, the pread does 256k.

The first read is SLOW. The read sizes should be 256k per
SIZE_DATA_BUFFERS right? We see 256K on the normal
files. 

Why is the process using read of the whole file, vs using a
quicker lseek across it?

Is there a way to tune the buffer size used when a sparse file
is detected? (ls size != du size)?

What about a better way to map out the holes?

Any tips on going faster would be apprecaited. (Multiplexing
provided some increase at 2 streams/drive, but recovery is
affected, we want really fast recovery).

Thanks!

Also - what is the real max of SIZE_DATA_BUFFERS? If we set to
4m, it does 4m in place of 256 above, and perf is better, with
plenty of ram the 512MB or so of shared memory is not an issue
at all (512 per tape drive).


4194304

+----------------------------------------------------------------------
|This was sent by johnj2063 AT gmail DOT com via Backup Central.
|Forward SPAM to abuse AT backupcentral DOT com.
+----------------------------------------------------------------------


_______________________________________________
Veritas-bu maillist  -  Veritas-bu AT mailman.eng.auburn DOT edu
http://mailman.eng.auburn.edu/mailman/listinfo/veritas-bu