some benchmark results
wensong at gnuchina.org
Sat May 11 15:15:47 BST 2002
On Mon, 6 May 2002, motse wrote:
> I benchmark the KTCPVS 0.0.7 and DRWS 0.0.0-alpha.
> benchmark environment
> 1. 6 real servers at back-end, each has 256MB ram, P-iii 1G and intel eepro
> 2. 3 clients run Webbench 4.1 with dual althon XP, 256MB DDR SDRAM and intel
> eepro 1000
> 3. The loadbalancer using p-iii 1G,256 MB ram and eepro100 or eepro1000
> 4. All devices are connected with one gigabit switch
> one web server with eepro 100 can provides 85~89 Mb/s web traffic
> KTCPVS 0.0.7 with intel eepro 100 ---> 1865 request/sec 89 Mb/sec
> DRWS 0.0.0 with intel eepro 100 ---> 8879 request/sec 472 Mb/sec
> KTCPVS 0.0.7 with intel eepro 1000 ---> 5643 requests/sec 298 Mb/sec
First, thanks for benchmarking the ktcpvs 0.0.7.
But, I don't think that it is reasonable to compare the numbers above
together, which may give the wrong impression. :) In LAN environment, it
is true that a web server with 100Mbps NIC can ship 85~90 Mbps traffic.
The round-trip time in the LAN is less than 1ms, but the average
round-trip time on the Internet is around 150ms, so in the Internet
environment, the real performance of a web server will drop significantly
(maybe 20Mbps or less), because request handling will occupy computing
resource for longer time. I hope that the KTCPVS doesn't consume too much
resource for each request handling, and I have to say that there are a big
space to improve the performance of KTCPVS, but for now we need make the
function of KTCPVS complete first, so that there will not be a fatal
exception that KTCPVS cannot handle in the real environment.
For comparing the performance of DRWS with KTCPVS, you can try to load
KTCPVS on your each Linux-based web server (like your masq module) and
IPVS on the front-end load balancer. :)
More information about the lvs-users