nanog mailing list archives
RE: 923Mbits/s across the ocean
From: Antony Antony <antony () phenome org>
Date: Sun, 9 Mar 2003 17:35:24 +0100
Les > The jumbo frames effectively increase the congestion avoidance additive Les > increase of the congestion avoidance phase of TCP by a factor of 6. Les > Thus after a congestion event, that reduces the window by a factor of 2, Jumbo frames helps a lot. Also NAPI in Linux 2.5.x will help a lot. The way most Linux (till 2.4.20) NIC drivers handle TX and RX interrupt moderation and SMP with multiple NICs are not very efficient. That really kill the performance of TCP over Long Fat Networks. Jumbo frames alleviate the problem very much. Poor x86 CPU is swamped with large number of interrupts when TCP send/receive burst of data about the size of congestion window, ~32Mbytes in our case. -antony
Current thread:
- Re: 923Mbits/s across the ocean, (continued)
- Re: 923Mbits/s across the ocean Iljitsch van Beijnum (Mar 09)
- Re: 923Mbits/s across the ocean David G. Andersen (Mar 09)
- Re: 923Mbits/s across the ocean Iljitsch van Beijnum (Mar 09)
- RE: 923Mbits/s across the ocean Cottrell, Les (Mar 08)
- RE: 923Mbits/s across the ocean Cottrell, Les (Mar 09)
- Re: 923Mbits/s across the ocean Richard A Steenbergen (Mar 09)
- Re: 923Mbits/s across the ocean Iljitsch van Beijnum (Mar 10)
- Re: 923Mbits/s across the ocean Richard A Steenbergen (Mar 10)
- Re: 923Mbits/s across the ocean Iljitsch van Beijnum (Mar 11)
- Re: 923Mbits/s across the ocean William Allen Simpson (Mar 15)
- Re: 923Mbits/s across the ocean Richard A Steenbergen (Mar 09)
- Re: 923Mbits/s across the ocean Stephen Sprunk (Mar 11)