On 04/29/10 03:54 PM, [email protected] wrote:
On Thu, Apr 29, 2010 at 03:19:22PM -0500, Shawn Walker wrote:
On 04/29/10 01:17 PM, [email protected] wrote:
The stats engine accounts for connection time as a separate value.
Have you verified that these changes still lead to appropriate
repository selection when faced with a large number of mirrors and
origins?
I don't quite follow; before we only called record_connection() if
conn_count> 0 and conn_time> 0 and that behaviour is unchanged for
http(s).
The only difference here really is that for some protocols,
conn_time could be 0 (such as file), so a "connection" is just
accessing the material.
The following comment implies that you've changed how the throughput
calculation is performed. I was curious if you've been able to verify
that this doesn't change the quality results in unexpected ways.
# Only performing this subtraction when the
# conn_time is 0 allows the first request in
# the pipeline to properly include connection
# time, etc. to initiate the transfer.
It depends on what time we want to count for the purpose of transfer speed.
It's been changed in the sense that previously, the first request's
performance in a pipelined sequence could be inflated because it
subtracted starttransfer_time so the time to connect, etc. wasn't
included although (it seems to me) it should be.
With a pipelined connection, the first transfer is generally slower than
the subsequent ones from what I can tell.
The numbers look "right" to me, but again, I may not have the correct
view of things.
But as far as I can tell, the repository selection behaviour is
still intact when multiple mirrors or origins are present although
I've only tested with a few, not "large numbers" (not even sure how
I would do that given there aren't many available). And yes, I've
looked at what happens when you mix 'http' and 'file' as well --
file is generally almost exclusively used and either gets a slightly
higher quality number of a significantly higher one depending on the
operation.
If there's a specific sort of test you're looking for, I can run
that if you tell me what it is.
There are about 5 mirrors on SWAN and about 3 origins. I'll send you
the instructions I used to test this in a separate e-mail. You'll need
to verify that you're getting the results that you expect. You may also
want to run this test on a machine on SWAN in sfbay and verify that the
results look right for that setup too.
Thanks, that will help me get a better set of results to compare.
Cheers,
-Shawn
_______________________________________________
pkg-discuss mailing list
[email protected]
http://mail.opensolaris.org/mailman/listinfo/pkg-discuss
