Bug 31940
| Summary: | [nge] ip queue length too short for >500kpps | ||
|---|---|---|---|
| Product: | Base System | Reporter: | fabien.thomas <fabien.thomas> |
| Component: | kern | Assignee: | freebsd-bugs (Nobody) <bugs> |
| Status: | Closed FIXED | ||
| Severity: | Affects Only Me | CC: | eugen |
| Priority: | Normal | ||
| Version: | Unspecified | ||
| Hardware: | Any | ||
| OS: | Any | ||
|
Description
fabien.thomas
2001-11-12 22:40:00 UTC
On Mon, Nov 12, 2001 at 02:39:27PM -0800, Fabien THOMAS wrote:
> when using SMC gigabit adapter (DP83820/DP83821) the link is slow 200Mb/s using netperf between 2 SMC adapter under FreeBSD and nothing more with jumbo frame and sometimes the card seems to renegociate the link (i've tried using different cable with no sucess)
Have you experimented with the options in the nge man page? There
are options to set a card in "master" mode, which sound like they
would help if you are using these cards without a switch. You
might also want to try the half-duplex and full-duplex options.
David.
yes i've tried in manual mode with link0 set on one side and in autoneg mode (the bandwidth problem seem to have no relationship with the media ramdon negociation) i've tried a netperf between two SMC gigabyte host (athlon 900 PCI32 and bi-athlon 1.2gb PCI64 (thunder K7)) and my max bandwidth is 300Mb/s thanks for your reply Fabien I am using 5.2.1, rather than 4.4, but I have found a bug which could be causing this. I have mailed freebsd-net but they don't seem interested. The problem is that the receive DMA ring is longer than the IP queue, such that if more packets are received in the interrupt than fit in the IP queue, packets are dropped. Since the default IP queue length is 50 packets, and the interrupt holdoff time is 100uS, you can only really expect to get 500kpps, which is not much for a gigabit ethernet card. You can fix by increasing the ip queue length constant. (make it >= 128) -- Luke State Changed From-To: open->feedback RELENG_7 defaults to net.isr.direct=1 - however, freebsd can do 10GigE even with net.isr.direct=0 State Changed From-To: feedback->open This is still an issue. Should change the default to 256 as on NetBSD. Responsible Changed From-To: freebsd-bugs->kmacy This is still an issue. Should change the default to 256 as on NetBSD. Responsible Changed From-To: kmacy->freebsd-net kmacy has asked for his PRs to be reassigned For bugs matching the following criteria: Status: In Progress Changed: (is less than) 2014-06-01 Reset to default assignee and clear in-progress tags. Mail being skipped Looks like nge(4)-specific issue. Now it seems to use 256 for both of transmit and receive ring length and interface sending queue. Is this PR still relevant? Feedback timeout. Also, requested change has already been made and problem's fixed in modern FreeBSD versions. |
