On Tue, 20 Nov 2007 14:00:10 -0600, Hal Merritt <[EMAIL PROTECTED]> wrote:
>Interesting question. My understanding is that this a feature of the >adapters. Therefore I ass-u-me that the data actually flows from the >host to the adapter back to the host. >... I don't think that is the case. The Miilicode/microcode is the same code that moves data to/from GBe OSAs but the adapters aren't involved. I seem to recall that sholder-tapping code had to be added for Hipersockets because there were no interrupts involved, but the actual transfer just involves buffer shuffling. Actually, buffer pointer shuffling. >If that is true, then one could SWAG the speed as something close to >theoretical gigabit. ... I'm guessing it probably faster unless that pseudo-interrupt slows things down a lot. >. .. However, you would still be constrained by IP and >Ethernet protocols, where a large percentage of the data flowing is >going to be TCP/IP headers and trailers along with gigabit (fast >Ethernet) headers and trailers. I think there are no GBe headers involved. "Large percentage" obviously depends on the data being sent. You get 40-50 bytes of TCP/IP headers > >A small MTU would set an upper limit on throughput the same way small >block sizes constrain other data paths. >... If you use MTU discovery thetre is no reason not to put a very large MTU size on Hipersockets. Ours (according to a NETSTAT DEV) is MTU Size: 65535. Pat O'Keefe ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [EMAIL PROTECTED] with the message: GET IBM-MAIN INFO Search the archives at http://bama.ua.edu/archives/ibm-main.html
