Question regarding connect reset errors on clients and possible load imbalance situation on Apache

[Yu-Ju Hong <yuru.hong@xxxxxxxxx>]

Hi,

I am new to Apache. I set up a distributed simulation using Simics (a full system simulator) to run Apache on the server system and Surge (Scalable URL generator) on the client system. Below is the configuration of the simulation.

Version: Apache httpd-2.2.13 (the latest version), compiled with mpm-worker

Simulated server system: Solaris 10 on Sparc

Simulated client system: Solaris 10 on Sparc

Simulator: Simics 3

The two systems communicate through a simulated Ethernet link.

1. On the simulated client system, I get some errors from time to time:

SURGEclient 0: Error on Read: connection reset by peer
Bad read of the first file
SURGEClient  0: Received SIGPIPE signal

Surge has more than 3000 clients running at the same time (in order to stress the system). Sometimes, the errors occur quite often. Is it possible that Apache wasn't set up correctly? Because Surge only has 25ms "thinktime" for each client thread, it is unlikely to be timed out by Apache.

2. I want to stress the simulated server system, which have 16 cores. However, by monitoring using "mpstat", only cpu 3 has constantly been used (0~12% idle time), while other cores have idle time around 75~85%. I have exhausted almost every possible way I know to stress more, but nothing has worked out so far. The methods I have tried are 1) increasing bandwidth up to 10Gb/s on both systems; 2) raising the number of Surge clients to more than 9000 clients; 3) Varying the maximum number of threads/processes of Apache (in httpd-mpm.conf). The situation remains the same, where only cpu 3 is heavily loaded while other cores just idle for most of the time. As far as I know, Apache should be more scalable than this. It is quite abnormal that only one cpu was stressed out. 

Hope I can get some suggestions on how to fix/check the above problems.Thanks!