In any case, it has been a resounding success. I'll summarise how things look at the moment; I'll do up a proper press release sometime later next month.
There's about 15,000 users sitting behind the single proxy cache server, with around 100mbit or so aggregate satellite IP bandwidth. The service uses a slightly modified FreeBSD-7 setup to support fully transparent HTTP interception (both client and server-side IP address spoofing) with a Cisco 3750 providing the WCCPv2 interception.
Tuning the FreeBSD stack (and Linux too, for those Linux people out there!) to effectively scale for satellite IP is no easy feat. It took a bit of time but I have quite a bit of experience in this area so the tuning was quite successful. The issue here is finding the right balance between throughput, scaling and link efficiency. A little bit of first year college mathematics helped me predict some decent settings and they work as expected.
The software is Lusca-HEAD (the very recent version as of this post) - this gives me all the useful Squid-2.7 features, stability and performance with my extras (twiddles for satellite IP stuff, TPROXY support, etc.)
The box itself is a dual dual-core AMD Opteron 270 at 2GHZ; 16gig RAM; Intel 1000/pro NIC, 3ware 9000 series SATA controller with 12 x 500gig 7200rpm disks of some sort. The disks are all mounted individually - no RAID at all. 10 disks are for storage; 1 for OS and 1 for logging.
The box pushes around 80 to 120mbit at peak with a byte hit rate between 20 and 40%. The request rate sits between 300 and 600 requests a second, sometimes peaking to 800 or more. This translates to traffic savings (saving a whole lot of money - satellite transponder space is expensive!) and much improved performance for clients.
It also handles between 10,000 and 20,000 concurrent connections with peaks over 40,000. Yes. 40,000 concurrent connections. I'm not making this up.
The cache size at the moment is around 2TB and 20,000,000 objects. I'm absolutely, positively not filling the disks to capacity for a whole lot of very good reasons. (Hint - don't do it.) I'll be happier to increase the storage to 4TB and beyond once I've deployed COSS for the small objects and tidied up some of the memory usage. The Lusca process is around 4 gigabytes at the present time and 75% that is the storage index and related bits.
Just for interests sake - out of the 20,000,000 objects, around 300,000 of them are larger than 256 kilobytes. The rest are small objects. It is quite scary actually how much of the cache directory is small objects.
I've included some preliminary windows update caching which is providing a 100% hit rate for the update files themselves. It's actually quite scary how simple it was to implement. Shame on you Microsoft for -almost- but not quite getting HTTP caching "right" in the windows updates.
All in all, the client in question is extremely happy about the support, installation and performance of the cache. There's a shortlist of items to do including Lusca improvements and reporting tools so the client can provide further information to his boss about how effective this all is.