> > questions:
> > 1. Does "platDur" affect the duration of fill phases?
> > On the client side, the real fill phases I observed are
> > 4.5 and 9.5 minutes respectively for platDur = 8 and 16 minutes.
>
> Looks like it does, indirectly. The fill phase doesn't exit until the
> working set size is frozen. Working set size depends on platDur:
>
> int wsc = int(peak_fill_rate * platDur); // number of objects in the
> WS
> working_set_cap(wsc / TheBench.client_host_count);
> Phase phFill = {
> ....
> wait_wss_freeze = yes; // will finish only if WSS is frozen
> };
>
[Xu, Zhonghui] good explaination. thanks.
>
> >
> > 2. On the server side, the fill phase I observed kept about 1.5 minutes
> > even for platDur =8, 16, 24 and 32 minutes, given that the pg files
> > on the both sides are same.
>
> I think the recently-added phase synchronization feature is not
> working perfectly. Some of the tests that I've run are out-of-phase
> as well.
>
> Note, it doesn't really affect the test. The client's log is
> perfectly valid and thats where most of the interesting statistics
> are. It does make the server logs difficult to analyze. You can't
> do some things like ask "what is the server's request rate during the top2
> phase."
>
[Xu, Zhonghui] we don't analyze server side log. but if the server
exits earlier than does the client, it causes trouble. we need to
give
an estimated, long enough time for the server to be alive with the
client.
>
> > What's interesting, if the server runs while the client does not,
> > the fill phase of the server takes more than 15 minutes even for
> > platDur = 8 minutes.
>
> I'm not sure I follow that. What sense does it make to run the server
> without a client?
>
> Duane W.
>
[Xu, Zhonghui] It doesn't make sense. I run this way just out of
curiousity to see if the client affects the server in sense of
"synchronization".
Thanks a lot.
Zhonghui
This archive was generated by hypermail 2b29 : Tue Jul 10 2001 - 12:00:15 MDT