[Q] incoming_icp_average

Hi,

I have the distinct feeling that some of my toplevel caches running 2.0.2
should handle more ICP requests, for instance from mgr:comm_incoming
during busy hours

Current incoming_icp_interval: 45
Current incoming_http_interval: 255

Histogram of events per incoming socket type
ICP Messages handled per comm_poll_icp_incoming() call:
        0 5228
        1 364954
        2 78726
        3 36497
        4 24394
        5 18327
        6 14873
        7 12375
        8 10537
        9 9022
       10 7749
       11 6769
       12 5970
       13 5187
       14 4587
       15 59819
HTTP Messages handled per comm_poll_http_incoming() call:
        0 1265
        1 128506
        2 34489
        3 18265
        4 11294
        5 7115
        6 4583
        7 2972
        8 1908
        9 1221
       10 3431

Quite visible is the sudden increase when managing 15 ICP requests. As 15
is the hardlimit here (INCOMING_ICP_MAX), I feel I need to tweak the
incoming_icp_average configuration option, currently at the default. Line
219 in comm_select.c didn't provide too much insight, except that this
parameter is the "target" mentioned in the comments at the start of file.
Now, should I set incoming_icp_average to 10, because that is the number
of requests I want to see processed on the average? Or do I have to set it
to 4, because I want to check the socket more often? Or is my server not
busy enough to worry about this?

Anyway, at peak times I am seeing up to 100 ICP/s reported processed,
which is 10/7 of the peak of my 1.1.x caches - currently my caches are in
transition, and some are already running 2.0.2 and some are still back at
1.NOVM.22. Still, it looks as if 3 % ICP go unanswered, and as if I am
hitting (yet another) wall.

Le deagh dhùrachd,
Dipl.-Ing. Jens-S. Vöckler (voeckler@rvs.uni-hannover.de)
Institute for Computer Networks and Distributed Systems
University of Hanover, Germany; +49 511 762 4726
Received on Wed Oct 21 1998 - 08:49:48 MDT