RE: CA performance

["Jeff Hill" <johill@lanl.gov>]

Steve,

I test the Channel Access performance of each new release of EPICS, but I
must admit that I am not good about keeping records of this. The program I
use most often to test performance is called "catime" it is provided with
both R3.14 and R3.13, but with R3.13 you must take additional steps to build
it as it isn't installed by default. 

My experience with "catime" has been that CA performance is currently
limited by the performance of the OS/driver/nic/network combination on 100
Mb networks. This conclusion is based on the fact that the CPU does not
saturate when the tests are running. This tends to change over time
depending on what LAN technology has been pushed out to the workstation.
Optimized builds *do* have a significant impact on efficiency.

Because, the IOC is a centralized resource the most significant performance
metric for EPICS is probably how efficiently records process and how
efficiently the resulting monitors are pushed out onto the LAN, but this
type of performance is difficult to measure with a CA client initiated test
such as "catime". We do periodically set up benchmark databases to gauge how
the performance of EPICS is changing over time.

Jeff

> -----Original Message-----
> From: Hunt Steven [mailto:steven.hunt@psi.ch]
> Sent: Monday, November 10, 2003 7:22 AM
> To: 'tech-talk@aps.anl.gov'
> Subject: CA performance
> 
> Has anyone carried out any channel access performance tests?, comparing
> performance for instance between:
> 
> Different versions of epics
> Different number of channels
> Different data lengths
> Different numbers of clients and servers
> 
> At the moment I am thinking of running a test varying the following;
> 
> Number of clients 1->8
> Number of servers 1->8
> Data length 1->1KB per channel
> Number of channels per server 1 -> 100'000
> Client connections per server channel 1->8
> get and put requests
> 3.13 and 3.14 clients
> Redhat Linux 7.3 and 9.0
> 
> All machines would be identical PCs running Linux, on an isolated
> network
> (one machine will be configured with two network interfaces to give
> access
> from the outside world). Servers would be 3.14. Network would be
> switched
> 100Mbit or Gbit Ethernet.
> 
> Any comments or suggestions?
> 
> Would anyone be interested in defining the tests that could be run
> and/or
> participating in the testing (you would be able to login to the
> machines).
> 
> regards
> Steve Hunt