A bar graph of our active backup sessions during the night is far from a Golden 
Rectangle. It's more like a big city skyline. We are backing up 510 clients 
now, but that number is increasing steadily. I am mulling over ways to smooth 
our session load. The hoped for benefit would be to make more effective use of 
available resources, and to delay the need for more TSM servers. We run TSM 
server version 6.2.2.2 on RHEL5 on a Dell R810 server. It's fibre connected to 
a Hitachi SAN.

Till now we have just aimed to average the number of associations within each 
schedule.That isn't effective since clients have widely different run times. 
Both of the following techniques use changes in Scheduling options to attempt 
to effect session smoothing. I would be glad to hear opinions on how well, or 
not, they might work.  

1) The first idea is to overlap Schedule Start Times. The randomization server 
option is there to avoid starting all backups in a Schedule together. That's a 
major knob for smoothing, but the highest setting for it is '50' which 
distributes actual start times over only the first half of a backup window. 
This allows clients to retry before the window expires. Since actual start 
times are spread over the first half of the window only, session count is 
necessarily higher in that half, which effectively de-smoothes the two halves 
of the window. Sawtooth City. 

We have few client communication errors, so we could set each Schedule's start 
time to be midway through the window of the previous Schedule. Then, the 
lightly used second half of the previous schedule would be filled with sessions 
from the first half of the overlapping schedule.  If Schedule Durations were an 
hour, Schedules would start every half-hour. There would be more defined 
Schedules, but fewer Associations per schedule. 

To be most effective, this technique may depend on the elapsed backup times of 
the clients in the overlapping schedules being more or less equal. So, it may 
need to be combined with the next technique which defines associations based on 
the experienced or expected elapsed run time of the client.

2) Another way to smooth sessions could be to make schedule Durations inversely 
proportional to the usual elapsed time of the clients associated with them 
rather than setting all Durations to an hour. About twenty-five percent of our 
backups finish in one minute or less. About fifty percent finish in five 
minutes or less. The easiest way to spread the good effect of all those light 
loads is to put them in a long window and let Randomization distribute their 
actual start times.  I am thinking of starting with a four hour duration for 
the clients with run times of less than one minute. 

Clients with long elapsed times would be associated with schedules of shorter 
duration so their start times could be more tightly controlled. Clients with 
intermediate elapsed times would be associated with schedules of intermediate 
duration. 

Since this technique depends on Randomization working as well with long 
Schedule Durations as with shorter, does anyone have experience with Durations 
of four hours or more? Did it go alright?

Any reactions to these ideas would be gratefully accepted.

With my thanks and best wishes,
Keith Arbogast
Indiana University