Comments on: Reverse Proxy Performance – Varnish vs. Squid (Part 2)

By: Torgeir

Torgeir — Tue, 02 Mar 2010 13:00:15 +0000

@Bryan

“I know both Facebook and Twitter use Varnish and I am sure they both handle and immense amount of traffic to their servers. I wish we could get that answer from them.”

There’s some information on these subjects in the list archives, for instance, John from Twitter posted about their setup here: http://projects.linpro.no/pipermail/varnish-dev/2009-February/000968.html

You’ll also want to look at Kristians blog, among other things he describes his 143k req/s setup: http://kristian.blog.linpro.no/2010/01/13/pushing-varnish-even-further/

By: Bryan Migliorisi

Bryan Migliorisi — Tue, 02 Feb 2010 12:35:15 +0000

Thanks Ronald!

I’ll check it out in Konqueror, though thats not the most common browser

By: Ronald Knudsen

Ronald Knudsen — Tue, 02 Feb 2010 12:32:50 +0000

Great post! I just suscribed to your RSS feed. Your site is kinda messy in my browser. I used Konqueror. Just to let you know.

By: Nikhil

Nikhil — Wed, 30 Dec 2009 05:54:14 +0000

Hey Brian,

Can u tell me how i can calculate the number of
— Maximum no.of Concurrent users during a time period(a day/a month)
— Maximum no.of concurrent sessions during a time period (a day/ a month)

Is it possible to calculate this from the access.logs?

Cud u tell me if u know any formula with which i can calculate the above.

Thanks
Nix

By: Durability, Scalability, Availability ·

Durability, Scalability, Availability · — Fri, 13 Nov 2009 01:28:02 +0000

[...] the fastest, which one is easier to setup or integrate with legacy apps etc. I’m certainly NOT trying to get into that! In fact I will skirt the issue entirely saying this: when the features are [...]

By: adrian chadd

adrian chadd — Sat, 07 Nov 2009 09:30:56 +0000

Did you publish the configuration somewhere? I seriously doubt seeing Squid’s concurrency falling off so sharply. I’ve been running Squid and Lusca in production with between 10 and 20,000 concurrent connections over a WAN with no noticable drop in connection timing.

You’re also not testing a “WAN” scenario well enough. You should introduce client and/or server side latency; packet loss; experiment with larger request and response bodies, etc, etc. You may find Varnish drops off quicker than you’d expect.

By: Bryan Migliorisi

Bryan Migliorisi — Wed, 28 Oct 2009 00:35:30 +0000

@Mark

Thanks for the pointer. I’ll rerun my tests using your recommendations.

By: Mark Nottingham

Mark Nottingham — Tue, 27 Oct 2009 20:47:26 +0000

http_load in -parallel mode tests how a server handles overload, not its capacity.

In other words, you’re showing (very effectively) how the servers handle a load greater than they’re able to, which isn’t necessarily at all related to how they’re able to handle well-sized loads.

While overload metrics are important, they’re not really giving any indication of how fast the server actually is; you’re only showing what happens once you exceed the servers’ capabilities.

To do that, you need to run a series of tests at different request rates, e.g., using autobench.

A few other metrics to look at: how different response sizes are handled, and proxy performance vs. cache performance.

By: Bryan Migliorisi

Bryan Migliorisi — Fri, 16 Oct 2009 16:52:55 +0000

@John

That is a good question. I know both Facebook and Twitter use Varnish and I am sure they both handle and immense amount of traffic to their servers. I wish we could get that answer from them.

By: John Connelly

John Connelly — Fri, 16 Oct 2009 16:33:50 +0000

I’d be curious to see a comparison of squid and varnish that deals with maximum concurrent connections. I currently use squid as a simple reverse proxy (no caching) to insulate apache from slow clients. It can handle ~5000 concurrent connections without problems. Unfortunately squid tops out around 3500 requests per second due to its high cpu usage and singe-process single-thread architecture.

Varnish, on the other hand, was able to do 13,000 requests per second in the same test setup, but when I tried it in a real-world environment it fell flat on its face with thousands of concurrent connections.

Anyone have any good tips for varnish running in simple reverse-proxy mode, and how to have it handle 10,000+ concurrent connections?