Performance and paralellism

[Iago77]

Nice to ask you again,

this time I would like to help me knowing the 'paralellism' term.

I've set parallellism server to 24. I'm working with Legato 6.1.3 and STKL700 (6 tapes). I've programmed a group whose backup time is often 1h and 30' with a lot of servers (clients) about 1h and 20 minutes (130GB). Last night I had only programmed a new group from a client of the group and another person changed paralellism in the client to 1. The backup time was 2 hours! although the save sets were about 50GB.

Is there someboy could help me to know how to adquire some information about the paralellism effect for backup?

Many thanks in advance!

 

[605]

This is not teasy to do without figures as there are a lot
of different NW 'parallelisms'.

The size of the save sets do not matter - the point is
whether you can keep the tape drive(s) streaming. If it is
a remote client with client save set=1 then only 1 save
stream can be opened, even if your server p. is set to a
higher value. Most likely, the data from this stream cannot deliver the data fast enough.

 

[eskolnik]

Well I am having the opposite issue my restores and clone process seems to be running extremely slow and I think it may be caused by intermixing too much data.

For one how could I tell if the tape drive is streaming or not?



My environment is
NW 6.1.3 on a Sun 280 R 2 GB RAM 36 GB hard drive (mirrored of course)
STK L40 Tape library with 3 SDLT 220 tape drives
About 90 clients (Window’s 2000, HP-UX, Sun Solaris, Linux (Redhat, Slackware) , and FreeBSD.

Recent Save Set Usage Summary:
439 today 349748 files 796/ss 85 GB 195 MB/ss 245 KB/file
461 7/29/03 327769 files 711/ss 165 GB 359 MB/ss 505 KB/file
276 7/28/03 475794 files 1724/ss 66 GB 240 MB/ss 139 KB/file
602 7/27/03 1834790 files 3048/ss 241 GB 400 MB/ss 131 KB/file
339 7/26/03 2169619 files 6400/ss 184 GB 543 MB/ss 85 KB/file
459 7/25/03 871366 files 1898/ss 280 GB 610 MB/ss 321 KB/file
467 7/24/03 376699 files 806/ss 66 GB 141 MB/ss 175 KB/file
3043 one week 6405785 files 2105/ss 1089 GB 358 MB/ss 170 KB/file





Ed Skolnik

 

[TDun]

A restore or saveset clone will run slowly IF the parallelism is set too high, causing the sessions to tape to go beyond the best performance for that tape drive barnd. There is a sliding curve for backup/recover performance on a tape drive. This alters with the drive type but an example (from DLT drives) would show 4 to be optimal (fast at both), where 1 gives fast recover but slow backup and 12 gives fast backup but slow recover. Between 4 and 8 seem to work best in my experience. Remember that this is per drive (target sessions).
Client and server parallelism need to be worked out to allow this to be achieved.

 

[eskolnik]

I thought change (Server parallelism 22 from 32) that was made to the backup server seems to have speeded things up quite considerably so I thought! The process of copying (cloning) save sets (client file systems or Windows drive's) has increased 10 fold. Yesterday the clone process was reading and writing at a rate of only about 2 MB/s , this morning I am seeing speeds from 10 - 33 MB/s, on one tape drive and only 3 MB/s when it rolled over to the next drive (1st tape became full).

My target sessions per drive was set at 33 so all the data would be sent to one drive at a time. With a target sessions set as 4 I will only be able to backup 12 (4 * number of tape drives) save sets at a time. I produce about 300 - 500 save sets a day, and have about 90 clients today .



Ed Skolnik

 

[TDun]

Unlike parallelism, target sessions is not a limit. If set to 4, this means it writes the first 4 streams to that tape. stream 5 goes to th next, as does 6,7,8. 9 would go to a third (and 10,11,12).
Stream 13 would go to the first, 14, to the 2nd, 15 to the 3rd, balancing at that point.

 

[eskolnik]

Actually I re-booted my Sun Solairs 9 NW 6.1.3 server and without any other changes cloning is woreking again!

There must be a memory leak or something that caused the problem.
Thanks to all who helped on this issue.

ed

Ed Skolnik