Hi Ernst
Thanks for your prompt reply. I will try once I get a better
understanding of what NAPI is? And is it useful to look into IRQ balance
as well?
Another question, what is the engineering limit for NISTNet traffic?
According to the calibration result on NISTnet webpage, it could be up
to 1Gbps. However, my testing results shows that the maximum supported is
80,000 packet/s×512byte×8=320Mbps.
Thanks
Stephen
-----Original Message-----
From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED]
Sent: 2007年12月4日 16:11
To: Stephen Wang
Cc: nistnet@antd.nist.gov
Subject: Re: [nistnet] [Help] - NISTNET occupy lots of soft interrupts
while traffic up
Stephen,
It could be that if youre system receives an interrupt for every packet
that is received and send. It is bound to choke at some point. You
could try to enable NAPI in the kernel. Which switches from interrupts
to polling when a certain level of interrupts is met. However you will
get some clustering of the send packets. i.e. the send inter packet
delay will vary somewhat more than with the interrupts.
Ernst
Dear NISTNET Supporter
I have been experiencing problems when testing NISTNET with Smartbit
6000C. Could you please help? Thanks in advance.
Test Platform
_____________<_____________
| eth1 | eth0
NISTNet server SmartBit 6000B
| eth2 | eth3
------------------------>----------------------
Smartbit(6000B) : 2¡Á1000M SX fiber
NISTNET Server: 2 x 3.2GHz CPU (Dual Core), 4G RAM, 2 ¡ÁHP NC373F PCI
Express Gigabits
Red Hat Linux Kernel: 2.6.13
NISTNET: 2.0.12c
IP£º
eth0 = 10.100.10.14/24
eth1 = 10.100.10.15/24
eth2 = 10.100.20.15/24
eth3 = 10.100.20.14/24
Test Strategy
Smartbit sends IP packet from eth0 to eth3; NISTNET server starts the IP
Forwarding; Meanwhile NISTNET injects delay.
1. Don¡¯t start NISTNET, drive through 800Mbps traffic from eth0, on
NISTNET server, with ¡°top¡± command see that ksoftirqd/0 process takes up
almost 99% CPU, about 0.01 % packet loss are seen on
eth3,
2. Start NISTNET, inject delay,jitter cnistnet ¨Ca 10.100.10.14
10.100.20.14 add new ¨Cdelay 15 5
The traffic exceeds 350Mbps, with ¡°top¡± command we see that SI (soft
interrupt) of CPU increasing dramatically. And even the server will go
down.
Enquiry
Why NISTNET will occupy so many SI resources of CPU when traffic is up?
Any ideas to improve?
Thanks
Stephen
--
This message has been scanned for viruses and is believed to be clean
_______________________________________________
nistnet mailing list
nistnet@antd.nist.gov
http://www-x.antd.nist.gov/mailman/listinfo/nistnet
--
This message has been scanned for viruses and is believed to be clean
--
This message has been scanned for viruses and is believed to be clean.