Ok, I found out the price of the barebones. Get a hold of something: :eek: :eek: $3,700. :eek: :eek: The reason apparently is because the case required to house this board needs some serious high quality power supplies. The board is about $1,700, so that makes the case at about $2,000. Appro engineers found this out the hard way on their 1U version of this system. This price puts my estimates earlier in this thread about $1000 too little for the barebones. nitro
would you care to DETAIL just what real-life performance gains are realized or are you talking thru your as? :-/
Caved in and getting it: http://www.tyan.com/products/html/tx46b4882.html Should arrive in about two weeks. Will populate it with two 2.0 Ghz 8xx series Opterons at first, and add two more Opterons later on. Eventually, I will populate it with dual core Opterons as their price comes down. I will try to give feedback on OS's compatibility and performance. This machine will probably force me to make the shift to multi-platform environments like QT since I cannot wait around for 64-bit windows. nitro
Will you try the XP 64-bit beta? http://www.microsoft.com/windowsxp/64bit/evaluation/upgrade.mspx Personally I will not spend the $$$ on a quad unless I first get a dual and successfully optimize my code for the dual. Who did you buy from?
Happy to see somebody talk some sense around here. Years ago, I used to run dual P3 Xeons to find out that programming for this is not very easy. I know better by now. Further, not having done the complete pricing analysis of the quad job, I would think thrice before spending one single $ for this. I have been very happy in the past by selecting my motherboards very conservatively. I have seen a lot of people crying about their experiences with gogo style boards. What are you going to do if something goes wrong with your four socket job, even if you populate it only with two processors. The tale about the cooling at a cost of about half the total should make one think twice. Isn´t reliability in general strongly linked to low temperature environments? All by all, only fools would want to take similar risks in the market. Of course many fools do.
Hi nononsense, Yeah, SMP isnât all that great anymore. One month ago, before I knew about NUMA I would have never considered buying a dual CPU machine⦠too expensive versus the performance. Multi threading will help some of my code, but not the important stuff because it's memory bound. If nothing else, a dual NUMA machine should be at worst equivalent to 2 singles, and at best perhaps 2x as fast as dual SMP systems if one maximizes use of the inter-CPU hypertransport and properly adjusts each threadâs affinity for CPU and memory bank (if possible). Same advantages apply for quads. Seems like Nitro got a solid system, with a solid PS, cooling, a good warranty and support. I cut corners by buying a pre-tested CPU+MB+MEM package, not a whole system. I did get one of the AMD recommended power supplies, the Antec True550 EPS12V. Knowing that he needs ultra-low latency computation on incoming market data, my real concern is does a quad like this provide any latency advantages over a fast single or dual? Can it achieve 20 GB/sec memory bandwidth across multiple hypertransport paths at a reasonable latency? Of course itâs always great for backtesting, where latency and I/O are less relevant.
I am not bound by the capacity to backtest super fast. The most important speed concern for me is how much computation I can achieve per "time slice" in realtime trading. My programs make use of lots of threads. If or when I need backtesting, I have my platform setup so that I can distribute the search over a cluster. nitro
Is it possible to control each thread's memory bank affinity, in addition to CPU affinity? Or are the 64 bit NUMA operating systems smart enough to allocate memory appropriately?
There was some talk about this on an earlier thread on ET. While it is certainly believable that it is possible to control the affinity of _PROCESSES_ per CPU, I really would be suprised if you could control thread or memory affinity. I am not even sure that you would want to, as the scheduler probably does a pretty good job based on what else is running. In my case, I would assign the TCP/IP stack to one CPU, and leave the other three up to the OS and that would suit me fine on a machine that is twice as fast as the speeding bullet I have now. For you and your design requirements, I suspect that the answer is no, but I am really vague on these details. nitro
My machine was supposed to ship friday. Now they are saying next two weeks. I have no idea what the hold up is. Everywhere I check no stock though. nitro