| Syren Baran 2007-05-30, 7:20 pm |
| toby schrieb:
>
> Well, that goes as far as 25,000 users. Is there a benchmarked system
> besides yaws that painlessly goes to 90,000 and beyond?
Oh sut up. Even the benchmark page states this simulates a DDOS attack,
not active users. Even estimating that every user on a web server will
only use 1KB/s (average) this would mean 90MB/s. Even in case of a GB
connection (in which case there would probably be server redundance,e.g.
round robin) this "benchmark" could only prove a bug or security flaw in
Apache. Post it to Apache or mitre as you wish.
On the other hand your benchmark actually details a far more serious
flaw in Jaws itself.
Why, tell me why, does Apache start out at 400KB/s on NFS while
Apache/local and Jaws/NFS do 800KB/s?
Oh, i got the answer.
800KB/s (given the stated 20KB packet size) require 40 packets/s, or, in
other words, 25 ms mean time. Given the fact that i assume tcp handshake
has been done (and HTTP/1.1 keep-alive is being used) i estimate only
two packets are necesarry, so a "ping Time" or network latency of 12.5
ms can be expected.
Now, if a web server checks (as it should) the mtime of a file over nfs
the same latency can be assumed (2 packets as well).
So, without knowing erlang or jaws, it guess its reasonable to assume
that jaws doesnt check for modifications of the files served.
Need to send it a HUP or restart it when you change a file in the server
directory, great, thanks, i think iŽll skip that one.
|