Bill Marquette wrote:
I saw a pretty significant performance boost on some of my IDS boxen
by putting the NICs on the same IRQ. There was also a tuning article
written quite some time ago (no idea about it's current day relevance)
that suggested the same. The IDS boxen have em(4) cards in them.
Interrupt load dropped tremendously and we gained back a lot of
dropped packets :) I had played around with recv timers and buffers
and was able to get a 10% performance improvement before changing the
cards to the same interrupt; I've since reverted those tunables as I
don't see a difference anymore.
Hey Bill-
Is IRQ sharing done in BIOS? I'm using 2 onboard em(4) NICs and a dual
port em(4) on a Supermicro 6023P-8:
em0 at pci3 dev 2 function 0 "Intel PRO/1000MT DP (82546EB)" rev 0x01:
irq 12, address: 00:30:48:2c:96:c6
em1 at pci3 dev 2 function 1 "Intel PRO/1000MT DP (82546EB)" rev 0x01:
irq 12, address: 00:30:48:2c:96:c7
em2 at pci5 dev 1 function 0 "Intel PRO/1000MT DP (82546EB)" rev 0x03:
irq 11, address: 00:04:23:a7:b5:08
em3 at pci5 dev 1 function 1 "Intel PRO/1000MT DP (82546EB)" rev 0x03:
irq 11, address: 00:04:23:a7:b5:09
We're pushing over 100Mb/s with a decent sized pf.conf, seeing 90%
interrupts during peak times. systat shows the interrupts split evenly
between an onboard port and one on the PCI card.
Here's our packet rate:
5 minute input rate 23209000 bits/sec, 15237 packets/sec; 5 minute
output rate 111684000 bits/sec, 16930 packets/sec
In addition to IRQ sharing, other things I wanted to explore:
* pf rules optimization
* further separating networks with additional NICs
* tuning the em(4) driver
* ensure PCI card is in 133mhz slot (pretty sure it is)
* 64-bit architectures (amd64 in particular)
any ideas ?
cheers,
sk
