FreeBSD performance tune up
Szymon Grabowski
I am running FreeBSD 4.3 on three servers:
a mail server (qmail 1.03 running on
512 MB RAM, 900 MHz box), a webserver
(1 GB of RAM, 900 MHz) and a MySQL database
server (1 GB of RAM, 1 GHz).
Are there any tweaks (i.e. kernel) that could
improve performance for high-volume servers?
On the side note, I am seeing the following on the
qmail server:
May 29 20:01:52 hermes /kernel: proc: table is full
May 29 20:01:52 hermes /kernel: proc: table is full
May 29 20:03:52 hermes last message repeated 354 times
What is causing this and how can I remedy?
Thank you for your help!
-- Simon
Doug Hardie
<szy...@profitstudio.com> wrote:
My servers required the following:
- increasing MAXUSERS to 64
- increasing NMBCLUSTERS to 4096
- set net.inet.tcp.delayed_ack=0
- set net.inet.ip.portrange.last=30000
The first 2 are kernel configuration parameters and require a new kernel
be built. I have heard some rumbling about being able to change them
dynamically, but have not looked into it. The last 2 can be changed
anytime with sysctl.
The MAXUSERS will increase the process table size and a few other tables
also. NMBCLUSTERS will increase the number of mbufs available for
tcp/ip. Delayed ack causes mbufs to be retained longer waiting for
delayed acks. Turning it off will reduce mbuf usage on high volume
servers (particularly web servers). Raising the port range provides more
ports to be available.
The above suggestions all came from the FreeBSD Handbook. There may have
been others that were not necessary for my situation.
Jason Anthony Mifsud
>
> In article <9f2k09$1tug$1...@news2.ipartners.pl>, "Szymon Grabowski"
> <szy...@profitstudio.com> wrote:
>
> > Hello,
> >
> > I am running FreeBSD 4.3 on three servers:
> > a mail server (qmail 1.03 running on
>
> - increasing MAXUSERS to 64
> - increasing NMBCLUSTERS to 4096
> - set net.inet.tcp.delayed_ack=0
> - set net.inet.ip.portrange.last=30000
> The MAXUSERS will increase the process table size and a few other tables
> also. NMBCLUSTERS will increase the number of mbufs available for
> tcp/ip. Delayed ack causes mbufs to be retained longer waiting for
> delayed acks. Turning it off will reduce mbuf usage on high volume
> servers (particularly web servers). Raising the port range provides more
> ports to be available.
Isn't there 65535 ports available as standard?
I'm not too sure, and am curious. :)
~JJ
Drew Lawson
Jason Anthony Mifsud <n...@no.com> writes:
>Doug Hardie wrote:
>>
>> - set net.inet.ip.portrange.last=30000
>
>
>Isn't there 65535 ports available as standard?
>
>I'm not too sure, and am curious. :)
On my 4.3-RELEASE system, sysctl tells me that is set to 5000.
The whole block is:
home:drew{14} sysctl net.inet.ip.portrange
net.inet.ip.portrange.lowfirst: 1023
net.inet.ip.portrange.lowlast: 600
net.inet.ip.portrange.first: 1024
net.inet.ip.portrange.last: 5000
net.inet.ip.portrange.hifirst: 49152
net.inet.ip.portrange.hilast: 65535
I don't know what these are based on, only that I haven't changed
them from what the kernal set as defaults.
--
Drew Lawson For it's not the fall, but landing,
dr...@furrfu.com That will alter your social standing
David Xu
Regards
--
David Xu
"Szymon Grabowski" <szy...@profitstudio.com> wrote in message
news:9f2k09$1tug$1...@news2.ipartners.pl...