Hi,

How much of a difference does having dual P3 CPUs over a single P3 CPU make in your web server? And *what* are the differences you notice when you are running a dual P3 web server over a single P3 CPU web server?

Does it make a difference in how many retail or reseller hosting accounts you'll put on a box? How much of a difference?

Thanks!

Louis

It all really depends on what the server runs. For things such as basic PHP and CGI apps youl never notice a difference in performance. But for more database intensive PHP/CGI scripts and ones with large overhead - youl most lickly see a nice difference. However the most difference comes with operations such as the generation of statisitcs from log files, and other CPU intensive stuff.

I have at times seen people upgrade to dual CPU systems only see notice a downgrade in performance (yes this does happen).
Its actually pretty common in bentchmark tests were Single CPU systems outperform dual CPU systems. It just all basicly comes down to what the server is running and how much overhead (CPU overhead) is being used. Dont upgrade to a Dual CPU system if you utilize allot of Memory resources and marginal CPU resources - that in itself is pointless.

Originally posted by rfxn
I have at times seen people upgrade to dual CPU systems only see notice a downgrade in performance (yes this does happen).


Are the two systems (before and after the upgrade) 100% identical?

Its actually pretty common in bentchmark tests were Single CPU systems outperform dual CPU systems.

Can you point me to what benchmark tests you are referring, and of course, the result of these tests? I have not seen a benchmark tests in which a single CPU outperforms Dual CPU.

cheers,
:beer:

Offhand i dont remember the software, it was awile ago. And yes the system specs were for the most part the same. This was a project i conducted while in school.

dual AMD 700mhz
512MB of SDRAM @ 133Mhz
1xmaxtor 20gb hard drive @ 7200RPM's
(had identicle system just utilizing a single CPU)

Most systems should ensure that the average process run on the system, meets the below criteria - before upgrading to Dual CPU:
1) The average process working set fits in the L1 or L2 cache.
2) The average process working set fits in main memory.
3) The average process does not requires a large amount of disk I/O.

E.G:
If you have a system with 300 sites on it and extensive disk i/o with only one HD then your going to lose peformance as the i/o operations to the hard drive must now be split between 2 CPU's. Not to mention adding a CPU cant make a hard drive perform i/o faster than its designed for.

The single CPU system had out performed the dual CPU system by almost 20% whilst performing extensive disk i/o operations.

But dektong, your write within reason - youl never find a single CPU system outperform a dual CPU system on intensive CPU operations (youl always get more flops out of the dual CPU system).

But essentialy, the CPU is only one critical system component. If your hard drive cant keep up with whats being thrown at it then your obvisouly going to have problems. I make it habbit to get 2 hard drives with dual CPU systems. As well Memory is another factor - if your system has low amounts of physical ram then the system as a whole can take a performance hit (including CPU operations).

Basicly you just have to balance hardware in a dual CPU system - make sure your hardware components can keep up pace with what the Dual CPU's can throw at it.

the above posters are mostly correct, but as RFXN mentioned, dual CPU is only 1 part of it.

the trick is to see if you can tune the server to the max of it's abilities. everytime I sit down to tune a windows server I have about 3 weeks worth of log files and performace log files to look at.

some of the things that I look for are:

I watch the disk I/O because small changes within that area can improve the system overall. ( nasty application is mysql for win2k, more I/O reads than anything else)

I watch the CPU utilization and the application that caused the spike. sometimes you'll find that those spikes are not random but are active because of 2 or 3 applications running at the right time together, so you have to figure out how to priorities the requests as they come in so that the server does not get to 100% alot.

I think I got it mostly figured out so that I can place about 120 - 150 active domains on a server or about 190 on a server that have some activity but not a great deal.


Mike

Hi Mike,

What kind of server (i.e. number and type of CPU, amount of memory, operating system, SCSI or IDE, etc) are you referring to with the number of active domains you mentioned?

Thanks a bunch,

Louis

Originally posted by stlouislouis
Hi Mike,

What kind of server (i.e. number and type of CPU, amount of memory, operating system, SCSI or IDE, etc) are you referring to with the number of active domains you mentioned?

Thanks a bunch,

Louis

bunch of questions that I never answer to the public directly. but here goes as general outlines

servers are win2k servers
cpu's 800 to 1200 mhz
memory = always have a min of a gig less than that your wasting your time in windows.
drive can be ide, then you have to pay attention to the I/O. The controller cards you buy ( spend the money right there on the controller cards right after memory ) have to be the best within your outlined budget.

SCSI same rules apply.

if I recall correctly ( or I just might not care to devulge things ) try to find controller cards that will offload from the cpu and have the largest read ahead cache you can find ), if you find these types of controllers then you can safely get a lower cpu mhz machine like 800 mhz and provide yourself with a nice cost savings.

I don't believe in dual CPU's systems as an important need of a host. Again I'm most likely wrong but the added cost for that CPU does not justify the gains that are produced. better to place another server on line and load balence the system.

Thier is a way that I can see the justificate for dual CPU ( or even quad CPU) servers . but the cost for the setup's and the requirements are huge. and requires that the host have a justifyable income under certain applications before even attempting the idea ( don't ask me what it is, I'm not interested in talking about it )

also look for a specific poster, I think he's Cpervira or something like that. The person has more knowledge in his left pinky than most people combined. he drops hints all the time about improving systems in general. fountain of knowledge. and I find that I'm always asking him something or atleast some clarification on certain points.

Mike

Hi Mike,

Thank you for sharing. Just so you know, I'm not trying to pry any trade secrets out of you or anyone else. I'm just interested in folk's real world experiences with their rack mount servers.

I frequent a few technical boards and forums where it's a common practice for folks to share quite detailed specs on the machine's they are talking about or asking for help with.


Thanks again and take care,

Louis

Would'nt a second CPU also provide redundancy? eg. one fails, the other will continue.

Joel Strellner

Originally posted by xerocity.com
Would'nt a second CPU also provide redundancy? eg. one fails, the other will continue.

Joel Strellner

depends on the motherboard, and the operating system, still you would be required to shut the server down unless the montherboard is equiped for a hot swap

Mike

Originally posted by stlouislouis
Hi Mike,

Thank you for sharing. Just so you know, I'm not trying to pry any trade secrets out of you or anyone else. I'm just interested in folk's real world experiences with their rack mount servers.

I frequent a few technical boards and forums where it's a common practice for folks to share quite detailed specs on the machine's they are talking about or asking for help with.


Thanks again and take care,

Louis

Nothing like that, it's that there are a lot of vindictive people within forums so I never want to give out ammo. good example of this is a pal of mine in germany. His DNS server's are some 550 mhz server. he got it used against him for over 3 months on a forum, to the point where he went out and spent an extra 400 for 2 more 550 ( used box also ) and made them redundant dns server with load balancing. Certain people know my config's right down to the CPU model number LOL.

but to the point, there are many ways to take a weak mhz server(s) and make then hum like F1 race car.

mike

I'd rather spend the additional money on purchasing one Higher End CPU i.e 1.26GHZ PIIIs than to buy 2 separate 800Mhz CPUs. Dual CPUs DO NOT yield double perfomance.

In addition, if you're running on Windows 2000 Server take note that there there is a large difference between a SQL Server Single CPU license vs. a Dual CPU license.

In addition, when running Dual CPU's there's always the additional (slight) risk of overheating assuming you have bad ventilation

Originally posted by mattan
I'd rather spend the additional money on purchasing one Higher End CPU i.e 1.26GHZ PIIIs than to buy 2 separate 800Mhz CPUs. Dual CPUs DO NOT yield double perfomance.

In addition, if you're running on Windows 2000 Server take note that there there is a large difference between a SQL Server Single CPU license vs. a Dual CPU license.

In addition, when running Dual CPU's there's always the additional (slight) risk of overheating assuming you have bad ventilation

your right on some points there. If you run MSSQL and Cold Fusion, you got to run the fastest you can get otherwise say hello to slow slow slow.

but if you have a defined client base and know there usage patterns I really can not see why you need to buy the 1600 - 1800 mhz cpu when you can go out and save 400 and buy a 1.26 pIII

for me atleast it's easyer to place another server on line and spread the load over a few servers, then try my hardest to have a nice balence of usage, and retune the servers. All my clients never have a problem when I redistribute them over the servers
because they know that they will be getting a nice little speed boost.

side note :
I rate each domain on the server on the basis of I/O load and cpu. When a new server comes on line, I then take a specic total of load and cpu from each server and place it on the new one.


in reference to the dual cpu license ( youch ) I myself rather split it into 2 servers and go out and get a second license to spread the I/O loads.

side note again:

instead of taking 1 x 40 hard drive, use 2 x 20 hard drives, why you have the databases run off 1 drive and the domains run off the other. Why, most controllers out there are good to handle the load split and can run extra hard without the problem.

mike

Hi Mike,

Thanks. From reading your posts, I infer that you prefer an add on card for SCSI rather than using the onboard SCSI that comes with some high end motherboards. Am I correct? If so, it *might* (wink) be because of the cache on a card over the lack of same with onboard SCSI, right?

Please feel free to PM me if you would care to answer -- but not in the public forum. I don't divulge info. shared in private, FWIW.

Anyway, if it's OK to ask, do you know if there is much of a difference in performance between using the onboard SCSI that comes with the high end Intel, Supermicro and Tyan dual P3 boards .vs an add on SCSI controller?

I've yet to build my first rackmount system -- or my first SCSI system for that matter.

Will likely build a 1U rather than a 2U. I know most 1U cases allow you to have a PCI card off a riser; but onboard is nice. Not sure how the extra performance a specific SCSI card might translate into how many extra domains a system can host.

I'm reviewing the motherboards and SCSI options right now. I don't have money to waste, so I'm wanting to pick the right components, that's all.


Thanks again & take care!

Louis

Originally posted by stlouislouis
Hi Mike,

Thanks. From reading your posts, I infer that you prefer an add on card for SCSI rather than using the onboard SCSI that comes with some high end motherboards. Am I correct? If so, it *might* (wink) be because of the cache on a card over the lack of same with onboard SCSI, right?

Please feel free to PM me if you would care to answer -- but not in the public forum. I don't divulge info. shared in private, FWIW.

Anyway, if it's OK to ask, do you know if there is much of a difference in performance between using the onboard SCSI that comes with the high end Intel, Supermicro and Tyan dual P3 boards .vs an add on SCSI controller?

I've yet to build my first rackmount system -- or my first SCSI system for that matter.

Will likely build a 1U rather than a 2U. I know most 1U cases allow you to have a PCI card off a riser; but onboard is nice. Not sure how the extra performance a specific SCSI card might translate into how many extra domains a system can host.
I'm reviewing the motherboards and SCSI options right now. I don't have money to waste, so I'm wanting to pick the right components, that's all.
Thanks again & take care!

Louis

Hi louis, yes I like off the motherboard processing, simple reasons is that every cpu cycle you remove from the motherboard you remove heat, and you can take advantage of the 32bit pci bus

1u cases are nice, make sure that you have the right motherboard that will let you get some good amount of ram with it. ( about 1 gigs is what you'll need if you run MSSQL and Cold Fusion and if you really want to host more then try to find 1 that will let you get to 4 gigs )

there is no exact rule for extra performance, the trick is knowing how to find the bottlenecks and clearing them up. once you got that solved you notice that sites run faster.

Now the last question is the clencher.
what is the value of the bottleneck in relation to hosting income.
if a bottle neckwhen cleared will reduce your overall average cpu load by 10%, that inturn means you should be able to load a few more sites within the server.

example : unix
500 domains, average cpu usage is at 60% ( running hot )
fix bottleneck and the cpu drops to 54 %, you should be able to load atleast another 20 domains on to the server ( 2/3's of the change).

20 x 6.95 = 139.00 extra revenue per server per month

now I would only be doing this sort of work when you get to 150+ clients, you got to get them first otherwise you risk spending money for no return.


but then again I could be completely wrong
http://slashdot.org/article.pl?sid=02/05/02/196212&mode=thread&tid=137
mike

Hi Mike,

Thanks again.


Take care and have a nice weekend!

Louis