diff options
Diffstat (limited to 'Documentation/networking')
| -rw-r--r-- | Documentation/networking/cxgb.txt | 72 |
1 files changed, 51 insertions, 21 deletions
diff --git a/Documentation/networking/cxgb.txt b/Documentation/networking/cxgb.txt index 9f2eb646c6f5..76324638626b 100644 --- a/Documentation/networking/cxgb.txt +++ b/Documentation/networking/cxgb.txt @@ -2,9 +2,9 @@ Driver Release Notes for Linux - Version 2.1.0 + Version 2.1.1 - March 8, 2005 + June 20, 2005 CONTENTS ======== @@ -21,8 +21,7 @@ INTRODUCTION This document describes the Linux driver for Chelsio 10Gb Ethernet Network Controller. This driver supports the Chelsio N210 NIC and is backward - compatible with the Chelsio N110 model 10Gb NICs. This driver supports AMD64 - and EM64T, and x86 systems. + compatible with the Chelsio N110 model 10Gb NICs. FEATURES @@ -121,23 +120,17 @@ PERFORMANCE Disabling SACK: sysctl -w net.ipv4.tcp_sack=0 - Setting TCP read buffers (min/default/max): - sysctl -w net.ipv4.tcp_rmem="10000000 10000000 10000000" - - Setting TCP write buffers (min/pressure/max): - sysctl -w net.ipv4.tcp_wmem="10000000 10000000 10000000" - - Setting TCP buffer space (min/pressure/max): - sysctl -w net.ipv4.tcp_mem="10000000 10000000 10000000" - - Setting large number of incoming connection requests (2.6.x only): + Setting large number of incoming connection requests: sysctl -w net.ipv4.tcp_max_syn_backlog=3000 Setting maximum receive socket buffer size: - sysctl -w net.core.rmem_max=524287 + sysctl -w net.core.rmem_max=1024000 Setting maximum send socket buffer size: - sysctl -w net.core.wmem_max=524287 + sysctl -w net.core.wmem_max=1024000 + + Set smp_affinity (on a multiprocessor system) to a single CPU: + echo 1 > /proc/irq/<interrupt_number>/smp_affinity Setting default receive socket buffer size: sysctl -w net.core.rmem_default=524287 @@ -151,8 +144,14 @@ PERFORMANCE Setting maximum backlog (# of unprocessed packets before kernel drops): sysctl -w net.core.netdev_max_backlog=300000 - Set smp_affinity (on a multiprocessor system) to a single CPU: - echo 00000001 > /proc/irq/<interrupt_number>/smp_affinity + Setting TCP read buffers (min/default/max): + sysctl -w net.ipv4.tcp_rmem="10000000 10000000 10000000" + + Setting TCP write buffers (min/pressure/max): + sysctl -w net.ipv4.tcp_wmem="10000000 10000000 10000000" + + Setting TCP buffer space (min/pressure/max): + sysctl -w net.ipv4.tcp_mem="10000000 10000000 10000000" TCP window size for single connections: The receive buffer (RX_WINDOW) size must be at least as large as the @@ -186,7 +185,7 @@ DRIVER MESSAGES may be found in /var/log/messages. Driver up: - Chelsio Network Driver - version 2.1.0 + Chelsio Network Driver - version 2.1.1 NIC detected: eth#: Chelsio N210 1x10GBaseX NIC (rev #), PCIX 133MHz/64-bit @@ -282,13 +281,44 @@ KNOWN ISSUES the number of outstanding transactions, via BIOS configuration programming of the PCI-X card, to the following: - Data Length (bytes): 2k - Total allowed outstanding transactions: 1 + Data Length (bytes): 1k + Total allowed outstanding transactions: 2 Please refer to AMD 8131-HT/PCI-X Errata 26310 Rev 3.08 August 2004, section 56, "133-MHz Mode Split Completion Data Corruption" for more details with this bug and workarounds suggested by AMD. + It may be possible to work outside AMD's recommended PCI-X settings, try + increasing the Data Length to 2k bytes for increased performance. If you + have issues with these settings, please revert to the "safe" settings + and duplicate the problem before submitting a bug or asking for support. + + NOTE: The default setting on most systems is 8 outstanding transactions + and 2k bytes data length. + + 4. On multiprocessor systems, it has been noted that an application which + is handling 10Gb networking can switch between CPUs causing degraded + and/or unstable performance. + + If running on an SMP system and taking performance measurements, it + is suggested you either run the latest netperf-2.4.0+ or use a binding + tool such as Tim Hockin's procstate utilities (runon) + <http://www.hockin.org/~thockin/procstate/>. + + Binding netserver and netperf (or other applications) to particular + CPUs will have a significant difference in performance measurements. + You may need to experiment which CPU to bind the application to in + order to achieve the best performance for your system. + + If you are developing an application designed for 10Gb networking, + please keep in mind you may want to look at kernel functions + sched_setaffinity & sched_getaffinity to bind your application. + + If you are just running user-space applications such as ftp, telnet, + etc., you may want to try the runon tool provided by Tim Hockin's + procstate utility. You could also try binding the interface to a + particular CPU: runon 0 ifup eth0 + SUPPORT ======= |