Bang for the Buck, comparing Celerons to Xeons (and PIIIs?)

George Bonser grep@shorelink.com
Sun, 12 Sep 1999 21:58:28 -0700 (PDT)


On 12-Sep-99 Alan DuBoff wrote:
>> 
>> Dual P-III 500:
>> 
>> real    2m10.220s
>> user    3m47.150s
>> sys     0m19.000s
> 
> George,
> 
> Your numbers are completely different than mine. You show a substantially
> lower real time than you do user/sys. As an example, I show the following for
> my dual Celeron:
> 
> real  4m23.739s
> user  4m11.700s
> sys   0m17.510s
> 
> My dual Xeons show:
> 
> real    3m38.532s
> user    3m22.530s
> sys     0m13.120s
> 
> Maybe I don't understand what real time is vs. user/sys, I thought real time
> should be about the same as user and sys combined.
> 
> At least it is on both of my systems. Why does your systems spend so much
> less
> time there?

You are seeing the benefit of SMP. By specifying -j4 I am able to use both
processors. The user time is the total time used but since it is concurrent on
two processors, the real time is less. Not much use in jacking it up past 4
with only two processors and on 2.3.x you can probably get an even better
user/real spread because of the finer grained locking. With 2.2, if one of the
processors is accessing a filesystem, the other must wait if it also wants to
access the same filesystem. With 2.3 the granularity is to the file level so as
long as I am accessing a different file, both may access the filesystem at the
same time.



> I was thinking of going to the 2.3.18 kernel on my dual Celeron box, since
> Linus seemed to imply at LinuxWorld that a lot of work had been put into the
> SMP code, and that it will scale well to 8 processors. That is important as
> Intel had announced 8 way support for Xeons. Up until now there really
> haven't
> been working 8-way systems other than NCR, or Seimens. At least that is my
> understanding, most everyone selling 4 ways are all using the Intel mobos
> (SC450NX and AD450NX).

There are still a few bugs in 2.3.18 ... one SCSI controller would not compile
for me but I will see how it runs. Look for Alan Cox to put out some 2.3.18ac
kernels while Linus is away. They will likely sync up in the pre-2.3.19 series
when he gets back.


> As I recall, you live very close to me, like down by Stevens Creek and
> Miller.
> I don't have any air conditioning either, but I have well cooled systems,
> they
> are both in SuperMicro 760a cases, my dual Xeon has a 400watt redundant
> cooling PS (2 fans on it) loaded up with CPU and drive fans, I'm fond of fans
> in systems...;-) My dual Celerons only have a 300watt PS, but it is also
> redundant cooling. These are nice cases...maybe a bit big for some people,
> but
> I like the ventialation on them.

I have additional case fans as well. One is controlled by the SuperMicro mobo
when the CPU temp reaches the threshold. Simple rule of thumb ... convection is
your friend. Draw air in at the bottom and exhaust it at the top. Case temp
should always be measured at the top center of the case if you are monitoring
the temp. Most mulitmeter manufacturers sell temperature probes now.


> 
> I don't understand your numbers above, I'm watching my Xeons compile right
> now
> and it shows consisten numbers, with real time being very close to the
> combination of user time and sys time.


Use make -j4 or maybe even higher and compare the numbers. I must say the
latest kernels are much more stable on SMP. I could kill <=2.2.10 at will with
high network activity. I can't anymore with 2.2.12 and pre2.2.13