Corsair is usually a very good and reputable memory manufacturer. Maybe you may contact Corsair Technical Support at 1-888-222-4346 for more info about the 8100 chipset and these or other memory modules. Here is the Official Corsair Memory Support Forum: www.houseofhelp.com/v2/forumdisplay.php?forumid=128 As for the ZipZoomFly return policy: http://www.zipzoomfly.com/jsp/Service/Return.jsp
yes, i know, that's why i'm asking. i've never seen you post any "insights" so i'm curious what practical trading advantage you believe this setup will get you.
Don't hold your breath. I do not talk about the way I trade nor the research I do save some of my research to a few trusted friends. nitro
Ok, I just got the two last Opterons that will give me a total of four. I also went and got the RAM for these CPUs. But before I install the CPUs and their corresponding RAM, I am going to fill all the memory banks for the two CPUs that are installed now and rerun Sandra tests one more time, just in case it actually matters that all four banks filled give more memory bandwidth. nitro
Ok, I filled all the banks on CPU 0 and CPU 1. I did not get any extra memory bandwidth out of it, but one interesting fact is that Sandra at least seems to give the same numbers from run to run of the memory bandwidth test. Go figure. Next is to install all the CPUs and all the RAM to each CPU. Then rerun Sandra tests and see if it matters. After this, it will end my Sandra tests and move me into real world trading tests. nitro
Ok got all CPUs in. Each CPU has two banks of memory so it is operating at 128 bits and NUMA is also on. As expected, the addition of CPUs added nothing to the memory benchmarks as they are not CPU bound. However, the big surprise is that the memory/cache test more than doubled in performance! Hmmm, that sort of makes sense, I guess. The CPU benchmark also exploded in performance. Now it destroys all dual Xeon systems, including a dual 3.8 Ghz dual Xeon system. That is one of the things the Tyan people told me, the more CPUs in an AMD the better they scale over Xeon systems. The CPU multimedia benchmark also exploded, though the dual 3.8 Xeon machines are just as capable in this case. Well, that leaves the Sandra tests for at least six months as there is nothing left to learn from them at this point. I think the next time I run Sandra tests is when I install the 64-bit version of Windows 2003 Server EE and the 64-bit version of Sandra. That should really begin to show the advantages of running the Tyan Quad Opteron machine and leave the dual Xeon machine far behind. We'll see... If I have time I will install the 64-bit version of FreeBSD and see how benchmarks run there, but that is a fun project I doubt I will have time to run at this point. I learned alot from this so far. The next "tests" are real with real money on the line. If I am able to I will post some results. Thanks to everyone that lent a helping hand. nitro
Nitro, Here are benchmarks for my system so far: Setup: Dual AMD Opteron 242, Tyan Thunder K8W S2885 motherboard 1GB OCZ Tech PC3200 CL2 ECC registered meory (4 x 256MB, 2 per CPU) Windows XP 64 Bit edition Tested with SiSoftware Sandra 2004.10.9.133 SP2B (IA32, No AMD64 code) config 1: bank_interleave=auto node_interleave=auto, no NUMA detected by Sandra (this mode interleaves memory addresses between CPUs on a per-page basis in order to normalize latencies) Mem bandwidth 4260/4128 MB/sec Cache & Mem Combined index 10250MB/sec Speed Factor 7.5 config 2: bank_interleave=auto node_interleave=disabled, NUMA detected by Sandra (memory addresses start with CPU0's memory, then CPU1's) Mem Bandwidth 6034/6043 MB/sec Cache & Mem Combined index 10788MB/sec Speed Factor 6.4 These numbers are fairly consistent between runs. Did your inconsistency between runs ever resolve itself when you added two more processors? It seems a low speed factor is desireable: âSpeed Factor: is a figure representing the speed differential between the CPU's cache and memory. The value is the ratio of the fastest cache (i.e. L1) bandwidth to the main memory bandwidth. (Lower is better, i.e. the memory is not very much slower than CPU's cache)â But now why is my memory bandwidth test giving 6 GB/sec while the combined index is 10 GB/sec? My memory bandwidth numbers should be around 10GB like yours. Maybe the Opteron 242s are a bottleneck at 1.6GHz. Most of the published benchmarks use 250s. Maybe my BIOS memory settings are sub optimal. Next time I have a chance to reboot Iâll write down my bios settings and post them to compare with yours. I haven't had much time to benchmark all combinations. Are your systems operating entirely on collocated servers? Can you trade unattended? I thought you were once opposed to this.
J, I only have time to answer this at this moment. I will get to the others later. I have no objection to colocated servers as long as you have the proper safeguards, for instance, two connections to the Internet to get to the colocated server in case one connection goes down. I also object to colocating a machine somewhere you could not physically drive to in half an hour in an emergency. But even in the case where you have taken all the precautions, my biggest objection is to allowing even systems that could run all by themselves without someone in "front" of the machine (thru Remote Desktop, etc) all day long. You also have to understand that not all of my stuff if 100% automated, some just presents ideas and I take them or don't. I have five systems, each which trade totally different instruments (options, futures, equities) and are based on totally different ideas. Some of these systems need to be colocated for one reason, while another system for another reason. But I am always monitoring what is going on, and if I go on vacation or at the dentist or whatever and I cannot get back to a work area that has a rock solid Internet connection, I may have my systems going gathering statistics, but not trading live on the markets. nitro
I agree with you mostly. My choices are partly due to the system designs I ended up with. These systems need to trade 24 hours/day. They see a lot of activity while I sleep, especially with EUR/USD. So there's no way I could supervise them all day long. Instead I rely on safeguards and small size/ small leverage. I check on the systems at least twice per day. Constant monitoring would only protect from unexpected data interruptions. Reliability has been getting better. I have to intervene about once every 40 days, usually at a small loss or profit due to the unexpected interruptions. Yes there is the slight risk of a significant loss, even at the low leverage I use, but thatâs not unlike many investment or long term trading risks anyway. Iâm in positions 70% to 90% of the day, and trade right through news and economic numbers.
Hi all, Since we are in a hardware thread I would like please a suggestion about getting a better configuration for my PC. I trade from overseas and my Pingplotter/GimmeIP software is always telling me in realtime I have about 180/210 ms of latency vs. the server of my broker and 15 hops to reach the host server. My connection is an DSL 640Kb with another DSL line as backup. The trading platform is Jtrader 2.83 or 5.x and I'm using a AMD XP 3200 with 512 Mb RAM Corsair 3500 C2 (2-3-3-7) in dual channel with Abit NF7-S v2.0; the FSB/CPU ratio is 1:1 at 414 Mhz. My goal is simply to have the price updates on DOME as fast as possible. I don't know whether CPU power, memory bandwidth, or memory chips speed DDR specs such as 2-2-2-5 or choosing a mb with a LAN on a CSA Intel as with 875P chipset, etc. have any influence or not. Since I will always be late vs a trader that connects in Chicago area by at least 170 ms, which are the optimization I can adopt with the LAN, CPU, mb, etc, for avoiding any further possible bottlenecks? Any direction for research is welcome. Thanks.