High Scalability Building bigger, faster, more reliable websites.

I've been using squid for different things for about 10 years now. Though it works for most cases, I've been asked to be cautious before putting it in high traffic sites. The person who mentioned this had mentioned that they had to restart squid every few hours under very heavy load.

I think that it's a kind of truism to say that as "load increased, performance eventually decreased." That's to be expected with most servers; once capacity is reached, performance often degrades exponentially as the server catches up processing all backlogged requests and still accepting new requests. I see this as an indicator it's time to scale out. I haven't done the math to calculate the cost effectiveness, but my gut instinct is that Squid is still one of the most cost effective ways to attack the problem as opposed to throwing more web servers into the pool. With its own built-in peer-to-peer caching network, Squid makes it far easier and more efficient to scale than web servers. This means that as you scale out the Squids, they can just request content from cache peers leaving the web servers free to handle new requests. Squids can handle enormous amounts of traffic well, but will get overwhelmed at a certain point; that's inevitable.

I am frankly surprised that YouTube had trouble serving it's thumbnail traffic from Squid. Google has been using Squid for it's Thumbnails on their image search, Orkut and likely on many other properties. To be fair, cache hydration is an issue when dealing with millions of small objects. I also wish Squid would use it's own object store that didn't involve storing each individual object as a file on disk.

As for Squid's stability under high-load, I am curious if it still suffers from problems like that. Squid is an actively maintained project, frequently releasing updates. It's come a long ways from v1. If someone is genuinely experiencing lockups under high-load, I really hope they get in touch with the maintainers of Squid so as to get to the root of the problem.

Has anyone tested Squid against Varnish? I've read some articles, and watched a couple of presentations by the author, and it looks pretty promising. The author claims that Varnish will provide significantly better performance and scalability when compared with squid.

I actually found that squid performs better when compiled with Tcmalloc (google's perftool).
Squid used to degrade over time (several months of uptime) with average load of 3000 requests/minute on my webserver. After compiling it with perftool, I'm no longer seeing the degradation.
In fact perftool is so good that I decided to compile it with MySQL and my server runs happily now.

fak3r:

With your squid vs varnish comparisons, are you working with constantly full caches ? Meaning, is the working set larger than the cache size ? Varnish only recently had LRU eviction put in, so it can handle a consistently churning cache. Or, are you testing with what can be served just from memory ?

Oyun: were your tests done with working sets that were larger than could fit in memory, or on disk? i.e., was varnish constantly evicting objects at the same time ? or were you just serving everything out of memory each time? I've yet to hear of a varnish install that was running with constantly full caches, which is why I'm sticking with squid.