Re: Slowdown on high-load machines with 3000 sockets

[Christian Schmid <webmaster@rapidforum.com>]

Nick Piggin wrote:
> Christian Schmid wrote:
> 
>> I already tried with 300 KB and even used a perl-hash as a 
>> horrible-slow buffer for a readahead-replacement. It still slowed down 
>> on the syswrite to the socket. Thats the strange thing.
>>
> 
> Do you have to use manual readahead though? What is the performance
> like if you just let the kernel do its own thing? The kernel's
> readahead provides things like automatic scaling and thrashing
> control, so if possible you should just stick to that.

Acutally, thats why I changed it to manual readahead. I thought the slowdown comes from some weird 
read-ahead settings. But with the manual read-ahead, this problem appears as well. Its still a 
phenomenon why this slowdown definetly disappears when I raise /proc/sys/vm/min_free_kbytes to 
1024000 (until it hits around 4500 sockets) but it seems not to work with smaller or bigger values....

> Also, can we get a testcase (ie. minimal compilable code) to reproduce
> this problem?

This is a complex code, thats why I really want to avoid this..... Feel free to ask for traces or 
anything you want to know though.

Thanks,
Chris