I reran my benchmarks with the latest "trunk" of Apache Traffic Server, to make sure we're not regressing. I also tweaked the number of worker threads a little, a gut feeling tells me that with Hyper Threading, our auto-scaling algorithm isn't optimal (and, it really isn't). Here are the latest numbers, running over a GigE network (two Linksys el-cheapo switches between clients and server)
3,160,237 fetches on 3,666 conns, 1,800 max parallell, 1.58012e+09 bytes in 30 seconds 500 mean bytes/fetch 105,341.10 fetches/sec, 5.26704e+07 bytes/sec msecs/connect: 1.46781 mean, 6.674 max, 0.093666 min msecs/first-response: 16.3333 mean, 615.34033 max, 0.121333 min
That is, 105k QPS (with keep-alive) for small objects, over the network. It's pushing 52MB of payload at this speed, but remember the average size is very small (500 bytes). My box is an Intel i7 920, Quad core.