Re: CPU usage of httpd+slowcgi

2020-08-17 Thread Kihaguru Gathura
Hi Alceu,

The Web applications in C, httpd+slowcgi+libpq are quite fast for very
heavy database querying.

Older dual core machines have limited memory up to 16 or 32 GB and moving
to newer machines like a T2000 or T4-1 is advantageous in the respect of
better memory capacity.

On the flip side, these newer machines have eight cores or more. The fewer
the better in this respect due to kernel locking considerations as advised.

What are reliable SPARC hardware alternatives to T4-1 that offer higher
memory capacity with fewer cores and low power consumption.

Kind regards.

Kihaguru



On Wednesday, July 29, 2020, Alceu Rodrigues de Freitas Junior <
glasswal...@yahoo.com.br> wrote:
> Hi Kihaguru,
>
> Oh, well, didn't see anybody asking, so... why are you guys still using
> a CGI if so much processing power is required?
>
> Basically anything will perform better nowadays, so unless the code is
> really complicated, you're looking for a solution at the wrong place.
>
> Em 27/07/2020 14:52, Jordan Geoghegan escreveu:
>>
>>
>> On 2020-07-24 03:16, Kihaguru Gathura wrote:
>>> Hi,
>>>
>>> Which of the following legacy CPU types is best suited for very busy web
>>> server httpd+slowcgi
>>>
>>> Niagara CPU Such as T2 - More parallel Threads and Low power per single
>>> thread
>>> Sparc64 CPU such as VI, VII - Fewer threads but more computing power per
>>> thread.
>>>
>>> How is multithreading utilization of httpd+slowcgi like?
>>>
>>> Kind regards,
>>>
>>> Kihaguru.
>>
>> Hi  Kihaguru,
>>
>> As with any computer, newer tends to be better with Moore's Law and all
>> that. On sparc64 most of the logical cores that are shown are really
>> just SMT pretending to be a bunch of cores. I have one machine that
>> claims 128 cores, but in reality, its just 16 cores with 8-way SMT.
>> sparc64 isn't renowned for its single core execution speed, so the
>> faster the better in that regard.
>>
>> In my experience with running OpenBSD on sparc64, the kernel biglock or
>> crypto became a bottleneck before other things did. (I've used T3 and T4
>> machines fairly extensively with OpenBSD). I've found that disk
>> activity, networking and/or TLS would bottleneck before httpd became a
>> bottleneck when I was running sparc64 web servers in production. If you
>> are running very heavy scripts/programs with slowcgi, then you're
>> results may be different.
>>
>> Things have likely improved dramatically in the past year or two with
>> all the work done on removing the biglock, but the moral of the story
>> remains, fewer, faster cores are likely to produce superior performance
>> to numerous low power cores.
>>
>> Regards,
>>
>> Jordan
>>
>
>


Re: CPU usage of httpd+slowcgi

2020-08-16 Thread Stuart Henderson
On 2020-08-16, Kihaguru Gathura  wrote:
> Hi,
>
> It depends on the workload. I'd have thought for most things the max
>> really usable at the moment is probably somewhere in the region of 4-8
>> cpu cores before kernel locking gets in the way too much.
>>
>> FWIW sparc64 ports builds are now done on T4 and they're really fast.
>> I think (but am not 100% sure) that this is carved into ldoms so the
>> number of cores visible to each OpenBSD instance is limited (so
>> contention between cores in the kernel is also limited).
>>
>>
> Up to how many cores will it be guaranteed that kernel locking 'will not
> get too much in the way' for hardware running a single instance of OpenBSD
> under heavy workload?
>
> Kind Regards,
>
> Kihaguru.
>

There are no guarantees.




Re: CPU usage of httpd+slowcgi

2020-08-16 Thread Kihaguru Gathura
Hi,

It depends on the workload. I'd have thought for most things the max
> really usable at the moment is probably somewhere in the region of 4-8
> cpu cores before kernel locking gets in the way too much.
>
> FWIW sparc64 ports builds are now done on T4 and they're really fast.
> I think (but am not 100% sure) that this is carved into ldoms so the
> number of cores visible to each OpenBSD instance is limited (so
> contention between cores in the kernel is also limited).
>
>
Up to how many cores will it be guaranteed that kernel locking 'will not
get too much in the way' for hardware running a single instance of OpenBSD
under heavy workload?

Kind Regards,

Kihaguru.


Re: CPU usage of httpd+slowcgi

2020-07-28 Thread Alceu Rodrigues de Freitas Junior
Hi Kihaguru,

Oh, well, didn't see anybody asking, so... why are you guys still using
a CGI if so much processing power is required?

Basically anything will perform better nowadays, so unless the code is
really complicated, you're looking for a solution at the wrong place.

Em 27/07/2020 14:52, Jordan Geoghegan escreveu:
> 
> 
> On 2020-07-24 03:16, Kihaguru Gathura wrote:
>> Hi,
>>
>> Which of the following legacy CPU types is best suited for very busy web
>> server httpd+slowcgi
>>
>> Niagara CPU Such as T2 - More parallel Threads and Low power per single
>> thread
>> Sparc64 CPU such as VI, VII - Fewer threads but more computing power per
>> thread.
>>
>> How is multithreading utilization of httpd+slowcgi like?
>>
>> Kind regards,
>>
>> Kihaguru.
> 
> Hi  Kihaguru,
> 
> As with any computer, newer tends to be better with Moore's Law and all
> that. On sparc64 most of the logical cores that are shown are really
> just SMT pretending to be a bunch of cores. I have one machine that
> claims 128 cores, but in reality, its just 16 cores with 8-way SMT.
> sparc64 isn't renowned for its single core execution speed, so the
> faster the better in that regard.
> 
> In my experience with running OpenBSD on sparc64, the kernel biglock or
> crypto became a bottleneck before other things did. (I've used T3 and T4
> machines fairly extensively with OpenBSD). I've found that disk
> activity, networking and/or TLS would bottleneck before httpd became a
> bottleneck when I was running sparc64 web servers in production. If you
> are running very heavy scripts/programs with slowcgi, then you're
> results may be different.
> 
> Things have likely improved dramatically in the past year or two with
> all the work done on removing the biglock, but the moral of the story
> remains, fewer, faster cores are likely to produce superior performance
> to numerous low power cores.
> 
> Regards,
> 
> Jordan
> 



Re: CPU usage of httpd+slowcgi

2020-07-27 Thread Jordan Geoghegan




On 2020-07-24 03:16, Kihaguru Gathura wrote:

Hi,

Which of the following legacy CPU types is best suited for very busy web
server httpd+slowcgi

Niagara CPU Such as T2 - More parallel Threads and Low power per single
thread
Sparc64 CPU such as VI, VII - Fewer threads but more computing power per
thread.

How is multithreading utilization of httpd+slowcgi like?

Kind regards,

Kihaguru.


Hi  Kihaguru,

As with any computer, newer tends to be better with Moore's Law and all 
that. On sparc64 most of the logical cores that are shown are really 
just SMT pretending to be a bunch of cores. I have one machine that 
claims 128 cores, but in reality, its just 16 cores with 8-way SMT. 
sparc64 isn't renowned for its single core execution speed, so the 
faster the better in that regard.


In my experience with running OpenBSD on sparc64, the kernel biglock or 
crypto became a bottleneck before other things did. (I've used T3 and T4 
machines fairly extensively with OpenBSD). I've found that disk 
activity, networking and/or TLS would bottleneck before httpd became a 
bottleneck when I was running sparc64 web servers in production. If you 
are running very heavy scripts/programs with slowcgi, then you're 
results may be different.


Things have likely improved dramatically in the past year or two with 
all the work done on removing the biglock, but the moral of the story 
remains, fewer, faster cores are likely to produce superior performance 
to numerous low power cores.


Regards,

Jordan



Re: CPU usage of httpd+slowcgi

2020-07-27 Thread Theo de Raadt
> Better to look for an M3000 or M4000 or as suggested for a T4-1

M3000 don't work.  The firmware locks up very badly.



Re: CPU usage of httpd+slowcgi

2020-07-27 Thread Claudio Jeker
On Mon, Jul 27, 2020 at 02:54:25PM +0100, Stuart Henderson wrote:
> Replying back on-list, I don't do support-type mails off-list, and other
> people know more about sparc64 hardware than me.
> 
> On 2020/07/26 22:38, Kihaguru Gathura wrote:
> > Hi Stuart,
> > 
> > For legacy, single-core CPU's such as Sparc64 V.
> > Would OpenBSD cope well with more number of CPU's or less as in previous 
> > case?
> > 
> > Example.
> > 
> > 2 CPU's (primepower 250) -> 4 CPU's (PrimePower 450) -> 8 CPU's(PrimePower 
> > 650) -> 16 CPU's
> > (PrimePower 850) -> 32 CPU's (Primepower 1500)
> 
> It depends on the workload. I'd have thought for most things the max
> really usable at the moment is probably somewhere in the region of 4-8
> cpu cores before kernel locking gets in the way too much.
> 
> FWIW sparc64 ports builds are now done on T4 and they're really fast.
> I think (but am not 100% sure) that this is carved into ldoms so the
> number of cores visible to each OpenBSD instance is limited (so
> contention between cores in the kernel is also limited).

The primepower 250 are decent and IIRC you can get dual core SPARC64-VI
CPUs for those. They use a fair amount of power. The bigger irons are fun
but honestly the weight and power consumption is just not worth it.
A primepower 250 is compareable with a fast v215. At least that is my
experience.

Better to look for an M3000 or M4000 or as suggested for a T4-1. Also make
sure you get good CPUs in them (esp. the M4000 comes with a few options).

-- 
:wq Claudio



Re: CPU usage of httpd+slowcgi

2020-07-27 Thread Stuart Henderson
Replying back on-list, I don't do support-type mails off-list, and other
people know more about sparc64 hardware than me.

On 2020/07/26 22:38, Kihaguru Gathura wrote:
> Hi Stuart,
> 
> For legacy, single-core CPU's such as Sparc64 V.
> Would OpenBSD cope well with more number of CPU's or less as in previous case?
> 
> Example.
> 
> 2 CPU's (primepower 250) -> 4 CPU's (PrimePower 450) -> 8 CPU's(PrimePower 
> 650) -> 16 CPU's
> (PrimePower 850) -> 32 CPU's (Primepower 1500)

It depends on the workload. I'd have thought for most things the max
really usable at the moment is probably somewhere in the region of 4-8
cpu cores before kernel locking gets in the way too much.

FWIW sparc64 ports builds are now done on T4 and they're really fast.
I think (but am not 100% sure) that this is carved into ldoms so the
number of cores visible to each OpenBSD instance is limited (so
contention between cores in the kernel is also limited).



Re: CPU usage of httpd+slowcgi

2020-07-25 Thread Stuart Henderson
On 2020-07-24, Kihaguru Gathura  wrote:
> Hi,
>
> Which of the following legacy CPU types is best suited for very busy web
> server httpd+slowcgi
>
> Niagara CPU Such as T2 - More parallel Threads and Low power per single
> thread
> Sparc64 CPU such as VI, VII - Fewer threads but more computing power per
> thread.
>
> How is multithreading utilization of httpd+slowcgi like?

Like most of the base system, httpd and slowcgi are not multithreaded,
also OpenBSD's scheduler doesn't really know about multithread cpus
(and even for full cpu cores, beyond a handful of cores OpenBSD doesn't
cope all that well).

Generally fewer faster cores is the way to go for OpenBSD.




CPU usage of httpd+slowcgi

2020-07-24 Thread Kihaguru Gathura
Hi,

Which of the following legacy CPU types is best suited for very busy web
server httpd+slowcgi

Niagara CPU Such as T2 - More parallel Threads and Low power per single
thread
Sparc64 CPU such as VI, VII - Fewer threads but more computing power per
thread.

How is multithreading utilization of httpd+slowcgi like?

Kind regards,

Kihaguru.