<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Strict//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-strict.dtd">
<html xmlns="http://www.w3.org/1999/xhtml"><head>
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8"/>
<style type="text/css">.mceResizeHandle {position: absolute;border: 1px solid black;background: #FFF;width: 5px;height: 5px;z-index: 10000}.mceResizeHandle:hover {background: #000}img[data-mce-selected] {outline: 1px solid black}img.mceClonedResizable, table.mceClonedResizable {position: absolute;outline: 1px dashed black;opacity: .5;z-index: 10000}
</style></head><body style=""><div>Hello, all. We are still trying to get to the bottom of our very poor performance on GbE connections with high (~=80ms) latency. We have noticed that on the production gateways (as opposed to test), software interrupts are through the roof. The major traffic was only 60 Mbps of UDP video to one site. Packet loss was horrendous even on barely utilized links and si was 100% on a single processor. pcrypt does not seem to help with the interrupts.</div>
<div> </div>
<div>We installed irqbalance and that helped quite a bit. We also played with /proc/sys/net/ipv4/xfrm4_gc_thresh and found an optimal setting at 2048000 but that only brought our end to end nuttcp throughput from 4 Mbps to 40 Mbps and eliminated the packet loss. We should be able to get at least 400 Mbps and I would really expect closer to 800 Mbps. We have proved we can exceed that on a single thread using UDP.</div>
<div> </div>
<div>What else can we do to get these software interrupts under control?</div>
<div> </div>
<div>kernel is 3.2.0-4-amd64 #1 SMP Debian 3.2.65-1+deb7u1 x86_64 GNU/Linux</div>
<div> </div>
<div>Thanks - John</div></body></html>