ZD Net tests, IRIX, and MAX_SOCKETS

[Jason Clary <jc...@futurefx.com>]

Hey guys.. i was reading this article in ZD Net's Internet Magazine
and I noticed that their unix server running IRIX 6.3 on an SGI O2
workstation with 180mhz R5000 processor, 128meg of ram, and 2gb harddrive
that its copy of Stronhold only managed to serve 100 requests/second
on average while on NT, MS IIS, NS Enterprise, and Commerce Builder
all managed over 500/s.. IIS was nearly 1000.

I was wondering if anyone knows if IRIX has a max socket limit in
the kernel like linux and SunOS do.  I believe most unix systems
are set to 256 socket maximum at boot time and can only be changed
via kernel rebuild.  I would expect the sockets to be taken up fully
in this case at around 100 requests/s in which case requests w ould
start waiting.

If this is the case, we should have a serious talk with ZDNet about that
as its giving ALL unix web servers a bad name, not just Apache/Stronghold
but all the others as well.  And its giving Unix a bad name.  I've
got a linux box running on a 486dx4100 that can do 100 requests per second
which makes this seem crazy.

Anyways, anyone running IRIX please take a look and see if its the same
sort of deal as on linux and SunOS where you recompile with 1024+ sockets
to get better performance on a net server.

NT, since the socket layer is not in the kernel, can allocate descriptors
on the fly... It also doesn't map them into its VFS like most unix systems
do and the descriptors aren't actualy interchangeable with file descriptors
like they are on unix so it doesn't suffer from this limit.  Although it
suffers in other ways because of the resource intensive nature of NT's
user interface.

If this is, indeed, the problem thats giving Apache such a bad benchmark
on their tests, it should be pointed out and corrected immediately.

Thanks

Jason

Re: ZD Net tests, IRIX, and MAX_SOCKETS

[Dean Gaudet <dg...@arctic.org>]

There are tunings that you have to perform in /var/sysgen/master.d/bsd on
5.3 and afaik on 6.2 as well -- so probably on 6.3.  But 6.2/6.3 don't
have the listen queue tuning parameter so I'm assuming SGI improved the
code such that it's either way high by default or it doesn't need tuning. 
That is where you would tune mbufs for example. 

A 128Mb 6.2 box has nproc = 576 by default, so it'd also run out of
processes unless systune was run on it.  They'd also have to systune maxup
so that the httpd uid could have more than 150 processes.  I doubt they
performed any of these tunings.

Which version of Apache?  Does keepalive figure into this? 

Aren't the other three servers multithreaded? 

Dean

On Mon, 20 Jan 1997, Jason Clary wrote:

> 
> Hey guys.. i was reading this article in ZD Net's Internet Magazine
> and I noticed that their unix server running IRIX 6.3 on an SGI O2
> workstation with 180mhz R5000 processor, 128meg of ram, and 2gb harddrive
> that its copy of Stronhold only managed to serve 100 requests/second
> on average while on NT, MS IIS, NS Enterprise, and Commerce Builder
> all managed over 500/s.. IIS was nearly 1000.
> 
> I was wondering if anyone knows if IRIX has a max socket limit in
> the kernel like linux and SunOS do.  I believe most unix systems
> are set to 256 socket maximum at boot time and can only be changed
> via kernel rebuild.  I would expect the sockets to be taken up fully
> in this case at around 100 requests/s in which case requests w ould
> start waiting.
> 
> If this is the case, we should have a serious talk with ZDNet about that
> as its giving ALL unix web servers a bad name, not just Apache/Stronghold
> but all the others as well.  And its giving Unix a bad name.  I've
> got a linux box running on a 486dx4100 that can do 100 requests per second
> which makes this seem crazy.
> 
> Anyways, anyone running IRIX please take a look and see if its the same
> sort of deal as on linux and SunOS where you recompile with 1024+ sockets
> to get better performance on a net server.
> 
> NT, since the socket layer is not in the kernel, can allocate descriptors
> on the fly... It also doesn't map them into its VFS like most unix systems
> do and the descriptors aren't actualy interchangeable with file descriptors
> like they are on unix so it doesn't suffer from this limit.  Although it
> suffers in other ways because of the resource intensive nature of NT's
> user interface.
> 
> If this is, indeed, the problem thats giving Apache such a bad benchmark
> on their tests, it should be pointed out and corrected immediately.
> 
> Thanks
> 
> Jason
> 

Re: ZD Net tests, IRIX, and MAX_SOCKETS

[Marc Slemko <ma...@znep.com>]

On Mon, 20 Jan 1997, Jason Clary wrote:

> 
> Hey guys.. i was reading this article in ZD Net's Internet Magazine
> and I noticed that their unix server running IRIX 6.3 on an SGI O2
> workstation with 180mhz R5000 processor, 128meg of ram, and 2gb harddrive
> that its copy of Stronhold only managed to serve 100 requests/second
> on average while on NT, MS IIS, NS Enterprise, and Commerce Builder
> all managed over 500/s.. IIS was nearly 1000.

They only tested one server under Unix?  That's lame.  Sure, they can
claim that they just wanted to test Unix as an alternative to NT so they
only needed one, etc. but it leaves an awful lot of room for
misconfiguration... 

> I was wondering if anyone knows if IRIX has a max socket limit in
> the kernel like linux and SunOS do.  I believe most unix systems
> are set to 256 socket maximum at boot time and can only be changed
> via kernel rebuild.  I would expect the sockets to be taken up fully
> in this case at around 100 requests/s in which case requests w ould
> start waiting.
> 
> If this is the case, we should have a serious talk with ZDNet about that
> as its giving ALL unix web servers a bad name, not just Apache/Stronghold
> but all the others as well.  And its giving Unix a bad name.  I've
> got a linux box running on a 486dx4100 that can do 100 requests per second
> which makes this seem crazy.

Are they doing SSL transfers though?  They are real pigs.  It is possible
(don't know, haven't looked) that the SSL part is slowing it down.

If they aren't doing SSL transfers, I agree 100/sec is way too low unless
they are doing some really odd transfers.

> 
> Anyways, anyone running IRIX please take a look and see if its the same
> sort of deal as on linux and SunOS where you recompile with 1024+ sockets
> to get better performance on a net server.
> 
> NT, since the socket layer is not in the kernel, can allocate descriptors
> on the fly... It also doesn't map them into its VFS like most unix systems
> do and the descriptors aren't actualy interchangeable with file descriptors
> like they are on unix so it doesn't suffer from this limit.  Although it
> suffers in other ways because of the resource intensive nature of NT's
> user interface.
> 
> If this is, indeed, the problem thats giving Apache such a bad benchmark
> on their tests, it should be pointed out and corrected immediately.
> 
> Thanks
> 
> Jason
>