Help:Performance measurements

Help:Performance measurements

Post by Nagaraj Nai » Wed, 27 Jan 1999 04:00:00

How do I get around with 'memory effects' (cache etc.) to get
accurate data on performance of programs ?



Help:Performance measurements

Post by Larry D Snyde » Thu, 28 Jan 1999 04:00:00

has anyone done any benchmarks on disk IO between Solaris 7, Linux and
FreeBSD?   More specifically with a 2940UW controller?

Also, how about network IO?



Help:Performance measurements

Post by Acacio Cruz » Fri, 29 Jan 1999 04:00:00

> has anyone done any benchmarks on disk IO between Solaris 7, Linux and
> FreeBSD?   More specifically with a 2940UW controller?

> Also, how about network IO?


I worked for an ISP in Summer98 and was involved in tuning their
proxy servers running squid on Linux.

Squid is a mammoth as far as network and disk usage is concerned.

To cut a long story short, we managed to reduce from 4 servers
serving a TOTAL of 12 reqs/sec on Linux to 3 servers serving 3 times
that much on EACH server running Solaris 2.6 and DiskSuite.

We did not change hardware. Just OS.

The major differences noted in our bechmarks were raw I/O ramdom access
speed (far superior in Solaris) and raw TCP/IP network performance.

Also, fine tuning of the IP stack made a major difference. Solaris
provides a very detailed access to its IP and TCP stacks.
For instance, changing a single parameter in the TCP stack "tcp_slow_start"
got us a 10 to 20 milisecond gain per HTTP request from Windows clients.
In our days, Windows rules and THAT was a major bonus.

Also, running DiskSuite to strip the filesystems had a huje impact on
squid performance. Some Linux sistems were using DPT controllers to do
the job but we found out that Solaris with DiskSuite had better performance
than Linux with DPT.

All testing was done using Intel EXCELENT motherboard's DK440LX.

BTW, I forgot to add that Solaris has a far superior SMP capability and
scales very well.

I don't want to start any religeous wars as I like ALL Unixes and
have experience with several.

Side notes: Solaris filesystem is derived from BSD so you should expect
similar filesystem performance from BSD. But DiskSuite is such a GREAT
feature that, if you'll  be running a fairly complex system, I must
recommend Solaris.

I'm curious to find out more about BSD network performance as I've been
getting very good reviews. I'll install it in my new box side by side
withj Solaris and run a few benchmarks.

I finished bechmarking ICMP performance on HP-UX and Solaris2.6 and the
results were clear: Solaris beat HP-UX 9 to 1 on inferior hardware !!


IT Consultant           Personal WWW site:
Dynamac Corp.

Version: PGPfreeware 5.5.3i for non-commercial use <>



1. performance measurement tools

I would like to characterize our product on Solaris 2.6 and 2.7
(and/or UW7). I am looking for a good set of public domain tools
and/or 3rd party products that would provide various counters at
cpu, memory usage, IO stats etc for an application (aggregate level)
and set of libraries (at function-by-function level). Our libraries are
a combination of C and C++ modules. Utility like vtune (from Intel)
for NT would be great.

Any pointers would be greatly appreciated.

Vipul Patel

2. login thru modem loses data

3. performance measurement for Pentium CPU

4. big IRQ lock removal, 2.5.27-G0

5. Performance Measurement on the WAN

6. Philips HCS-80 dumb terminal

7. Performance measurements for Linux on PowerPC?

8. Kermit Not Disconnecting -- Why?

9. Process Group Performance Measurement

10. Looking for performance measurement tools

11. performance measurements across 8 platforms: atoi()/sprintf("%d")

12. Performance measurement via vmstat puzzling

13. performance measurements on unix-systems (V.3 680X0)