On Thu, 25 Apr 2002, mukesh agrawal wrote:
> What I am doing is using wait_xact_lmt to approximate what might
> happen if the network is too slow to keep up with the specified
> request rate. So the robot would adapt its load to the network
> capacity by dropping requests when the network can't keep up. It's
> not obvious to me whether or not the dropped embedded object
> requests should be counted against the target request rate.
The actual (wait_xact_lmt related) question you have to answer is what
happens _after_ the "too slow" network recovers. The current model
tries to recover immediately. If you make the modifications you
mentioned, the request rate may increase more gradually. By
wait_xact_lmt definition, while the network is too slow, all new
transactions must be dropped.
Now let me try to address your specific needs rather than discuss a
possibly irrelevant piece of code. :)
If I understand your needs correctly, the right way to model your
environment is to introduce "stalled" transaction timeouts to
terminate a connection that had not seen any I/O activity for the
given (configurable) time. This way, you can model a more realistic
user behavior when the user cancels download after waiting for too
long.
The model can be further improved by introducing timeouts for active
transactions. That would be a bit more difficult because a user will
most likely want to cancel transactions that receive just a little bit
of data every so often (too slow) as opposed to transactions that
receive a lot of data for a long time (too large). See what I am
getting at?
Note that none of the above improvements are related to wait_xact_lmt.
Moreover, if the improvements are enabled, I would argue that
wait_xact_lmt should NOT be used if you want to simulate human
behavior. Again, wait_xact_lmt purpose is to prevent RAM exhaustion
and polyclt process swapping. The limit is not meant to model
realistic human behavior. It might simulate behavior of some automated
agents, but you probably do not care about that.
Finally, take a look at Rptmstat feature / PGL type. It might be of
interest in your situation:
http://www.web-polygraph.org/docs/reference/pgl/types.html#type:docs/reference/pgl/types/Rptmstat
HTH,
Alex.
This archive was generated by hypermail 2b29 : Mon Feb 06 2006 - 12:00:22 MST