Date: Sat, 01 Mar 1997 08:28:29 -0800 From: David Greenman <dg@root.com> To: hackers@freebsd.org Subject: Re: overview of socket hashing changes, plus new patch Message-ID: <199703011628.IAA05199@root.com>
next in thread | raw e-mail | index | archive | help
------- Forwarded Message
To: "David S. Miller" <davem@jenolan.rutgers.edu>
cc: netdev@roxanne.nuclecu.unam.mx
Subject: Re: overview of socket hashing changes, plus new patch
In-reply-to: Your message of "Fri, 28 Feb 1997 21:26:57 EST."
<199703010226.VAA08431@jenolan.caipgeneral>
From: David Greenman <dg@root.com>
Reply-To: dg@root.com
- --------
A couple of comments:
> FreeBSD - Single hash for entire TCP socket
> space. Listening sockets pollute
> the established connections.
I didn't see a significant amount of overhead due to this when I
implemented the PCB hashing in FreeBSD. Almost all of the overhead was
in the per-packet linear list searches...so creating a seperate hash
list for listening/wildcard PCBs didn't seem worth it at the time.
> Hash
> function involves a mod instruction
> because the table size is a prime.
> (Some minutae, I have approximately 50 or so samples from some
> of the largest ftp and web servers that I know about (some
> of these samples have 5,000 or more connections), listing
> the socket connections during high load. A comparison of the
> current bound hash function under Linux and the one
> under FreeBSD shows them to distribute about the same, if not
> better for Linux (in particular Linux is better on machines
> which run many clients as well as large single port servers)
> the prime size of the hash does not seem to help FreeBSD and
> they eat the multi-cycle modulo instrucion on top of it all,
> so stupid)
My analysis showed that an add or shift+add didn't provide adequate
distribution. I actually went to a lot of trouble to write the prime
number based hashing code because of this. If you've found a good way to
get similarly good distribution, I'm certainly interested.
> 2) Timers:
>
> FreeBSD - Every 500ms, and every 200ms, the
> _ENTIRE_ TCP socket list is walked,
> yes in it's entirety. How the fuck
> does this scale?
It doesn't, and it's one of the biggest problems with the code. It is,
however, a second order problem (7/second compared to 1000's/second in
the first order linear list search per packet case). I plan to re-visit
this eventually, but it's not been very high priority.
- -DG
David Greenman
Core-team/Principal Architect, The FreeBSD Project
------- End of Forwarded Message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199703011628.IAA05199>