On Sun, 27 Oct 2002, Richard Clarke wrote:

>     Before I embark on a day exploring the pros and cons of today's
> cleaning/compression tools, I wondered if any of you could give me some
> feedback about your own experiences within the context of medium/large scale
> web sites/applications (E-Toys etc).
> 
> Is it too presumtious to expect that many users now have high speed
> connections and tools that clean and/or compress html are of a small benefit
> anymore? That is not to mention the proliferation of "pretty" websites with
> 90% graphics.

We'are using mod_deflate on rambler.ru about 1.5 year. It's one
of the biggest Russian portals and search engines. We use gzipping
on caching and accelerating mod_accel frontends but not on
mod_perl backends.

We use conservative mod_deflate settings i.e. we do not compress
responses to requests that go through any proxy servers.
These settings allow us to save about 5-10% of all bandwidth.

Of course benefit of compressing depends to clients and content.
Many Russian users have slow connections and even they have fast enough
links these links usually shared so their resulting bandwith is low.

Graphics is usually good cacheble but texts is not always.
If it's your case then gzipping allows you to save bandwidth.


I see 3 some trades off of compression using.

1. Perfomance. mod_deflate has capabilty to check system idle time
(on FreeBSD only) and to disable gzipping if idle time would be less
then specified. Anyway I never saw less then 30% idle time on out frontends.
It seems that modern CPUs can easy gzip several small enough (30K) responses.

2. Memory. zlib uses about 300K for compressing. It seems to me more
important resource then CPU. Additional 300K is big enough to lightweight
frontend so mod_deflate has directive to disable gzipping if there are
more Apache childs then specified - this allows to avoid intensive swapping
when number of Apache childs increases for some reason.

3. Browser bugs. But it seems that modern browsers have not serious
gzipping bugs.


Igor Sysoev
http://sysoev.ru

Reply via email to