Skip site navigation (1)Skip section navigation (2)
Date:      Sun, 22 Mar 1998 11:50:51 -0800 (PST)
From:      Simon Shapiro <shimon@simon-shapiro.org>
To:        Andrew Reilly <reilly@zeta.org.au>
Cc:        current@FreeBSD.ORG, taob@nbc.netcom.ca
Subject:   Re: Worldstone Continued...
Message-ID:  <XFMail.980322115051.shimon@simon-shapiro.org>
In-Reply-To: <199803220833.TAA14256@gurney.reilly.home>

next in thread | previous in thread | raw e-mail | index | archive | help

On 22-Mar-98 Andrew Reilly wrote:
> On 21 Mar, Simon Shapiro wrote:
>> My very humble opinion is that we are barking up the wrong tree.  Disk
>> I/O
>> is not the limiting factor here.  I am looking at disk bandwidth
>> utilization of less than 20%.  From idle observation, it appears that
>> parsing Makefiles is 4-5 times the amount of time it takes to actually
>> compile anything.
>> 
>> The  curious could take the kernel directory and write a simple script
>> to
>> compile it without make at all.  Tell us the results.
> 
> For the kernel, that's nothing more than the output of make -n.  I
> think that building a script for make buildworld could be a little
> trickier.
> 
> I'm a little surprised that you think that make itself is so
> expensive.

Simple observation;  make clean && make depend && make -j8 in the kernel.
Observe the relative time it takes make to prepare what to do, and the time
it takes to actually compile it.

Another Observation;  I can increase disk bandwidth for large compilations
(make world, etc.) almost linearly with only minimal gain in overall time.
I compiled on RAID-5, RAID-0, with stripe width up to 8 drives, with source
and objects on separate arrays.  WRITE cache is set to write-back, the
system can demonstrate 5-19 MB/Sec for random I/O on small blocks, up to
1740 disk ops/sec.  Make buildworld and make -j8buildworld are resulting in
200, 280 disk I/O ops/sec, respectively.  CPU utilization is 80% user 70%
system (figure this one out :-).  I conclude from these numbers that the
system is not starved for I/O.

There could be one of several explanations:

* Make, which is low-I/O consumer, takes inordinate amount of time
* Compilations are CPU intensive, more than we think
* The current SCSI layer is somehow slow

I wish I had the time to port the DPT driver to CAM.  This will give us an
answer.

----------


Sincerely Yours, 

Simon Shapiro
Shimon@Simon-Shapiro.ORG                      Voice:   503.799.2313

To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-current" in the body of the message



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?XFMail.980322115051.shimon>