| TIP: Click on subject to list as thread! | ANSI |
| echo: | |
|---|---|
| to: | |
| from: | |
| date: | |
| subject: | Re: Crappy Windows 2000/XP UDP performance |
From: Mike '/m' I play with media here at home. At work I play with databases (business data in the databases, no media or rich content). I stumbled upon gigabit (here at home) because the price had come down a lot, and I needed some way to move large files (3-5GB) from my workstation to my media server reasonably quickly. One other comment, don't expect to get much more than 12-13MB/s on gigabit unless/until you move the gigabit NIC off the PCI bus and onto something faster. A lot of the motherboards that offer gigabit nowadays put the NIC on the northbridge bus (I think it's northbridge, might be southbridge, though). This gets around the PCI bottleneck. The newer versions of the PCI bus should also help. Some rough empirical throughput measurements from my setup: 100mbps ethernet - 7-8MB/s throughput 1000mbps ethernet on PCI bus, regular frames - 10MB/s 1000mbps ethernet on PCI bus, jumbo frames - 12MB/s 1000mbps ethernet off PCI bus, jumbo frames - 13MB/s At which point I believe I hit the bus bottleneck in the server I was copying to. I haven't dug into that yet. >How is life different when you slap on a. a heavy load of >"classical," small-data SQL-ish requests, vs. when the data follows the >current imagebase fad,... Gigabit ethernet gets the data into the memory of the server more quickly. How it is handled from there depends upon many things. I agree with your implication - this is a performance tuning challenge. :-) /m On Sat, 21 May 2005 13:56:47 -0400, "Frank Haber" wrote: >MikeM, do you do media work, or rich databases, or both? I've seen you make >some extremely knowledgable replies to questions about media arcana, as well >as to ones about hi-perf networking. > >Which brings me to some idle questions I have as a generalist. > >Gigabit Ether Throughput: > >o In 1993, with the best (?EISA) NICS and the fastest 486 then available, you >could load up 10Base to 20% and get 70-80% of theoretical max throughput. I >don't remember the loaded figures, but with heavy loads you could at least >expect 20-40% of theoretical, and response times (unless you were unlucky) >were still in the hard disk ballpark. CPU usage was a problem. > >o In 1999, with 100Base, still running 1500 packets, with PCI dual NICs on the >server, the figures were more like 60% best-case, lightly loaded, and the same >old 33%, loaded. CPU usage and streaming were both problems. > >o In 2005, would you care to make a comparison, with Gigabit, and big, >RAM-heavy servers? How is life different when you slap on a. a heavy load of >"classical," small-data SQL-ish requests, vs. when the data follows the >current imagebase fad, where every damnfool handwritten scrawl gets stored at >high res and indexed umpty-leben ways, including the squashed insects, the >form grid, and the coffee stains? --- BBBS/NT v4.01 Flag-5* Origin: Barktopia BBS Site http://HarborWebs.com:8081 (1:379/45) SEEN-BY: 633/267 270 5030/786 @PATH: 379/45 1 106/2000 633/267 |
|
| SOURCE: echomail via fidonet.ozzmosis.com | |
Email questions or comments to sysop@ipingthereforeiam.com
All parts of this website painstakingly hand-crafted in the U.S.A.!
IPTIA BBS/MUD/Terminal/Game Server List, © 2025 IPTIA Consulting™.