I'm doing my own realtime support for Asterisk, in an attempt to make it scale. Asterisk is nice software, but straight out-of-CVS, the performance for high volume (say, over 20,000 clients) sucks. There are also other inconviences with using a file-based store to determine how to route calls. Mainly, it's inflexible and hard to achieve high-perf when everything is based on a large .conf file. Not to mention that Asterisk uses linked lists for everything so looking up any user is an O(N) op (and parsing the users file is O(N*N) by default!). So, I'm going to put my own logic as a replacement for some of the critical parts.

One of my concerns was performance. Since I'll have multiple Asterisk clusters banging on my .NET code (via SOAP), I wanted to ensure the whole end-to-end process was fast enough.

I used gSOAP to create the C code on Linux. gSOAP is seriously nice. At least an order of magnitude easier to use than I expected any SOAP library that works on Linux would be.

I created a simple test. I made a database with phone numbers and codecs. The idea is that when an incoming call comes in, Asterisk will use my code to SOAP over to my Windows machines, get the data, and then go on its merry way.

My Asterisk machine is a P4 2.4GHz, 512MB RAM (but, I have a Gnome session running on it). My Windows XP machine (I tested against my desktop) is a P4 3GHz HT, 1.5GB RAM. I'm running ASP.NET 2 Beta 1 and SQL Server 2000.

The test program consists of a loop (count 5000) that generates a random number, then uses gSOAP to ask for the codec for that number. Simple, tight.

The results on Linux are particularly impressive. Each instance of the test app only used a max of 4% processor, and under 1MB of RAM. The bottleneck was definately inside ASP.NET. To simulate more load from other machines in a cluster, I ran 1, 2, and 4 instances of the test program. Also note that background tasks on the XP machine used up about 10% of the CPU.

Results:
 Single process (5000 total requests): 
  Total time:                   18 seconds (0.0036s/request)
  Requests per second:   277
  ASP.NET/IIS CPU:          30%
  SQL Server CPU:            4%

 Dual process (10,000 total requests):
  Total time:                   23 seconds (0.0048s/request)
  Requests per second:   384
  ASP.NET/IIS CPU:          60%
  SQL Server CPU:            7%

 Quad process (20,000 total requests):
  Total time:                   42 seconds (0.0052s/request)
  Requests per second:   476
  ASP.NET/IIS CPU:          80%
  SQL Server CPU:           10%

These results are encouraging enough that I'm not worried of the performance impact of using SOAP with Asterisk. My target was to have a response in less than 0.1 seconds. Although, anything under 0.5s would be quite unnoticable to a client. Even in tests with more threads, my single request response time was always way under 0.1 seconds.

Also, as far as I know, Whidbey Beta 2 (the version I'll go live with) makes some performance improvements. And also, IIS6 on Windows 2003 is much faster than IIS5.1 on XP. At any rate, a single proc desktop machine serving 476 RPS? That's pretty damn good perf if you ask me!