diff options
Diffstat (limited to '2005/flow-accounting-ols2005/OLS2005/grossman/grossman-abstract.tex')
-rw-r--r-- | 2005/flow-accounting-ols2005/OLS2005/grossman/grossman-abstract.tex | 31 |
1 files changed, 31 insertions, 0 deletions
diff --git a/2005/flow-accounting-ols2005/OLS2005/grossman/grossman-abstract.tex b/2005/flow-accounting-ols2005/OLS2005/grossman/grossman-abstract.tex new file mode 100644 index 0000000..dbbeb54 --- /dev/null +++ b/2005/flow-accounting-ols2005/OLS2005/grossman/grossman-abstract.tex @@ -0,0 +1,31 @@ + +% Registration Large Receive Offload implementation in +% Neterion 10GbE Ethernet driver +% Leonid Grossman (leonid@neterion.com) + +The benefits of Transmit Side Offload (TSO) +implementation in Ethernet ASICs and device +drivers are well known. TSO is a \textit{de facto} +standard in 2.6 Linux kernel and provides +significant reduction in \%cpu utilization, +especially with 1500 MTU. On a cpu-bound +system, these cycles translate into dramatic +throughput increase. Unlike TOE, stateless +offloads do not break the Linux stack and do +not introduce security and support issues. +Stateless offload benefits are especially +apparent at 10 Gigabit rates. 10GbE sender +with TSO hardware support uses a fraction of a +single cpu to run at line rate, leaving plenty +of cycles for applications. On the receiver +side, the Linux stack presently does not have +a stateless offload similar to TSO. Receiver +\%cpu typically becomes a bottleneck that +prevents 10GbE adapters from reaching line +rate with 1500 mtu. Neterion hw/sw Large +Receive Offload (LRO) solution was designed to +address this bottleneck and further reduce TCP +processing overhead on the receiver. Both +design and performance results will be +presented. + |