Quote from: slush on February 12, 2011, 10:40:56 AM

Many users ask me why I don't want their better hosting, why I don't set up fee for only CPU miners (which are traffic-ineffective), why I don't finish new protocol and so.

Please understand, that I'm on holiday and I'm writing from Internet cafe. I cannot deploy any big change, I cannot move system to another location. This is temporary hot fix for current system and I don't want to set up fees for forever.

I'm not exactly sure where the statement that CPU miners are traffic ineffective comes from, nor where/why the community's popular view is such. The 4way SSE algorithm puts one of my (VMWare ESX hosted so the GPU in it isn't usable) 6 core AMD Phenom's at 18khash/sec, and at a far lower $ cost than the spare cycles I run my GPU miners on with a pair of 5870's. That single CPU (which is a 2.7ghz X6, 1045T) does 18k/sec at a similar cost to a 5870. GPUs are definitely more efficient, but a well written algo making use of CPU SIMD instructions will come well within shooting distance of all but the latest highest end cards. My dual quad core opteron running at 2.0Ghz does another 14k/sec, granted that older box isn't anywhere near as efficient either, but the CPU miners could be faster, the SSE algo may not be the fasted algorithm in all cases (it IS way faster than the c algo or crypto++).

Now if you mean because each of the threads from each of the miners has to request work individually instead of as a team, I can understand that, but you're still only doing around 400hits/sec (if the getworks/sec is to be any indication of the hitrate) so I definitely agree that the amt of traffic exchanged for each getwork is probably quite a bit. One thing I noticed is the diablo GPU miner does not back off immediately when it gets an error. It retries immediately quite a few times before finally backing off and the site does 500 errors right at the time the network adds a new block, which is definitely causing more traffic for you. Probably doubling your actual traffic just sending out 500 errors instead of forcing clients to wait a moment.

Heck with blocking a bunch of people with auth errors you might make things worse because they'll all start hammering.

I know it's going to take quite some work to make BPM fly though, and it won't be easy.