Hi.

	It looks like T1000 has different motherboard than T2000.
Should there be any differences in performance of those two servers
with the same amount of RAM (ie. 4GB) exlcuding disk of course.

I'm thinking web servers.

We did production tests with T2000 and numbers are really amazing,
and it makes much more sense to put T1000s instead of T2000 for
web servers (price/performance and 1U vs. 2U). But I'm concerned
if for example T1000 has slower memory access, or smaller bandwith
to network adapters, or....

???

-- 
                                                Robert Milkowski
                                                rmilkowskiASDASD@wp-sa.pl
                                                http://milek.blogspot.com

I don't think the T1000 will have slower memory access, because the
memory access depends on the T1 processor which is the same in both
models.

The only thing to worry about with the T1000 is that do you really need
32Gb of memory?
Do you really need mirrored hdd in your server?
Do you need redundent power supply?
Do you need lots of expansion cards?

Since you want to use this for a web server expansion probably not an
issue, so if the answer to any of the questions above is yes then your
going to have to but the T2000, if its no to all the questions then the
T1000 seems the box for you.

Robert Milkowski <rmilkowski-NO-SPAM@wp-sa.pl> writes:

>	It looks like T1000 has different motherboard than T2000.
>Should there be any differences in performance of those two servers
>with the same amount of RAM (ie. 4GB) exlcuding disk of course.

Where did you get this information?

(My Hardware handbook doesn't list the T1000 yet).

Is this not a 1.0GHz vs 1.2GHz thing?

Casper
-- 
Expressed in this posting are my opinions.  They are in no way related
to opinions held by my employer, Sun Microsystems.
Statements on Sun products included here are not gospel and may
be fiction rather than truth.

Casper H.S. Dik <Casper.Dik@sun.com> wrote:
> Robert Milkowski <rmilkowski-NO-SPAM@wp-sa.pl> writes:
> 
> >       It looks like T1000 has different motherboard than T2000.
> >Should there be any differences in performance of those two servers
> >with the same amount of RAM (ie. 4GB) exlcuding disk of course.
> 
> Where did you get this information?
> 
> (My Hardware handbook doesn't list the T1000 yet).
> 
> Is this not a 1.0GHz vs 1.2GHz thing?

I've got T2000 and it looks different than T1000 on pictures (I mean motherboards).
First thing to notice is T2000 has 16 memory slots while T1000 has only
8 memory slots. So isn't throughput to memory in T1000 smaller than in T2000?
The memory throutput in Niagara boxes is one of the main factor when it
comes to performance.
Why there're ano SPECweb05 for T1000 (only T2000) ('coz it is not available yet?)?

Compare these two pictures of motherboards - they definitely look different.

http://www.sun.com/servers/coolthreads/t2000/gallery.jsp?img=5
http://www.sun.com/servers/coolthreads/t1000/gallery.jsp?img=4

As T2000 looks excellent in our tests we would really buy T1000s but
we are not sure if we could expect the same performance or not.



-- 
                                                Robert Milkowski
                                                rmilkowskiASDSADA@wp-sa.pl
                                                http://milek.blogspot.com

kangcool <kangcool2002@yahoo.co.uk> wrote:
> I don't think the T1000 will have slower memory access, because the
> memory access depends on the T1 processor which is the same in both
> models.

Latency - yep.
Throughput could be better in T2000 due to more DIMMS.
And that's what bothers me (and memory throutput is one of the
main factors in niagara performance).

> The only thing to worry about with the T1000 is that do you really need
> 32Gb of memory?
> Do you really need mirrored hdd in your server?
> Do you need redundent power supply?
> Do you need lots of expansion cards?

I know answers to all these questions.
All I need to know is there any performance difference
between T1000 and T2000 due do less memory slots in T1000
and/or possibly due to differen IO chips (GbE) with the same
amount of RAM? (excluding disk).

I guess I will have to check it myself but it's gonna take time.


Another interesting thing - why Sun doesn't offer T1000 with 1.2GHz
CPUs?

-- 
                                                Robert Milkowski
                                                rmilkowskiXCZC@wp-sa.pl
                                                http://milek.blogspot.com

Robert Milkowski <rmilkowski-NO-SPAM@wp-sa.pl> wrote:
> kangcool <kangcool2002@yahoo.co.uk> wrote:
>> I don't think the T1000 will have slower memory access, because the
>> memory access depends on the T1 processor which is the same in both
>> models.
> 
> Latency - yep.
> Throughput could be better in T2000 due to more DIMMS.
> And that's what bothers me (and memory throutput is one of the
> main factors in niagara performance).

I think you're right - as far as I know, 
T1000 only uses two memory controllers instead of four,
essentially halving the peak memory bandwidth compared to T2000.
Depends on your workload, 
two T1000s may perform better than a single T2000 
- it all depends on which bottlenect you're hitting.
If your application can extract all bandwidth of T2000,
your app won't run faster on two T1000s than one T2000.
However, if your app is I/O bound on T1000, 
two T1000s will perform better than a single T2000.
If your app is CPU-bound, then two T1000 will most likely 
do better than a single T2000.

> I know answers to all these questions.
> All I need to know is there any performance difference
> between T1000 and T2000 due do less memory slots in T1000
> and/or possibly due to differen IO chips (GbE) with the same
> amount of RAM? (excluding disk).

I don't think there's any significant difference in I/O or networking.

> I guess I will have to check it myself but it's gonna take time.

Yeap. The best way to determine is to run your workload on those systems.
I guess there's a try-and-buy program for T2k (though I'm not sure whether 
T1k is available for that as well).

> Another interesting thing - why Sun doesn't offer T1000 with 1.2GHz
> CPUs?

I'm glad I don't know the answer - I can have fun with speculation :)
My guess is that it's the result of combination of a few factor - 
CPU vs bandwidth balancing, the power consumption (and cooling), 
pricing and target market, 
simplicity of having less number of configurations, etc.
-- 
#pragma ident "Seongbae Park, compiler, http://blogs.sun.com/seongbae/"

Seongbae Park <Seongbae.Park@Sun.COM> wrote:
....
>> I know answers to all these questions.
>> All I need to know is there any performance difference
>> between T1000 and T2000 due do less memory slots in T1000
>> and/or possibly due to differen IO chips (GbE) with the same
>> amount of RAM? (excluding disk).
> 
> I don't think there's any significant difference in I/O or networking.

Well, looks like I was wrong. There are non-trivial differences,
and depending on how far you're pushing the I/O (both disk and network), 
you may see some difference.

See:
http://www.sun.com/servers/coolthreads/coolthreads_architecture_wp.pdf
for more detail.
-- 
#pragma ident "Seongbae Park, compiler, http://blogs.sun.com/seongbae/"

> Well, looks like I was wrong. There are non-trivial differences,

Anyone know if it's possible to insert a FC HBA into the T1000 and hook
it up to a SAN ? This would be the ideal way to overcome the disk
issues, and is something I'm investigating - although no one seems to
know (there are no HBAs listed as options for the T1000 for instance).

-Mark

Well you can kick your reseller into giving you one to see how it goes,
you will not get the 60 days but you should have enough time to see if
it meets your needs.

I can show you at least 5 opteron boards which can look different, be
perform the same.

>From the look of it the T1 has 4 memory controllers which can take 4
dimms each, so the T1000 only has 8 dimms so probably only uses 2
controllers.

It also depends what you do with the ram, from the look of it you will
see the better performance, if you using 1gb sticks then you will see
no difference, if you pack a T2000 out with 256mb sticks you will get
much better than T1000 since you can use all the memory controllers and
interleaving

Seongbae Park <Seongbae.Park@sun.com> wrote:
> Robert Milkowski <rmilkowski-NO-SPAM@wp-sa.pl> wrote:
> > kangcool <kangcool2002@yahoo.co.uk> wrote:
> >> I don't think the T1000 will have slower memory access, because the
> >> memory access depends on the T1 processor which is the same in both
> >> models.
> > 
> > Latency - yep.
> > Throughput could be better in T2000 due to more DIMMS.
> > And that's what bothers me (and memory throutput is one of the
> > main factors in niagara performance).
> 
> I think you're right - as far as I know, 
> T1000 only uses two memory controllers instead of four,
> essentially halving the peak memory bandwidth compared to T2000.
> Depends on your workload, 
> two T1000s may perform better than a single T2000 
> - it all depends on which bottlenect you're hitting.
> If your application can extract all bandwidth of T2000,
> your app won't run faster on two T1000s than one T2000.

This is bad news... :( T2000 is not that cheap - taking
performance factor into accounnt cost is similars to x86.
(I know - heat, rack units, etc.).

I wonder how can I measure how much memory bandwith
is consumed on our T2000 during tests - any clue?

I also wonder if I can pull out half of the ram from two
memory banks in T2000 so it will work with only two memory
controllers and then see what performance I get.
?


> > I guess I will have to check it myself but it's gonna take time.
> 
> Yeap. The best way to determine is to run your workload on those systems.
> I guess there's a try-and-buy program for T2k (though I'm not sure whether 
> T1k is available for that as well).

I've got our own T2000 and another one in try-and-by program.
Now we're going to get T1000 for testing (or buy one). Unfortunatelly
it will take another 4-6 weeks at least until it will arive.



Thank you for your very informative answers.

-- 
                                                Robert Milkowski
                                                rmilkowskiASD@wp-sa.pl
                                                http://milek.blogspot.com

Robert Milkowski <rmilkowski-NO-SPAM@wp-sa.pl> wrote:
> I wonder how can I measure how much memory bandwith
> is consumed on our T2000 during tests - any clue?
> 

bash-3.00# while [ 1 ]; do busstat -w dram0,pic0=mem_reads,pic1=mem_writes -w dram1,pic0=mem_reads,pic1=mem_writes -w dram2,pic0=mem_reads,pic1=mem_writes -w dram3,pic0=mem_reads,pic1=mem_writes -n 10 1|awk 'BEGIN {r=0;w=0;} {r+=$4;w+=$6} END {r=r*64/(1024*1024*10);w=w*64/(1024*1024*10);print "read(MB/s): "r; print "write(MB/s): "w;print}'; done
read(MB/s): 91.9781
write(MB/s): 89.8039

read(MB/s): 278.95
write(MB/s): 131.335

^C
bash-3.00#


Server with small load.


-- 
                                                Robert Milkowski
                                                rmilkowskiXSAQW@wp-sa.pl
                                                http://milek.blogspot.com