Re: RFC: doubling IO_BUFSIZE

[Pádraig Brady]

On 05/24/2014 05:21 PM, Jim Meyering wrote:
> On Sat, May 24, 2014 at 1:59 AM, Pádraig Brady <address@hidden> wrote:
>> On 05/24/2014 06:32 AM, Jim Meyering wrote:
>>> It looks like it makes sense to double IO_BUFSIZE once again.
>>> What do you think?
>>
>> +1
>>
>> Significant enough to bump up I think,
>> and we never saw regressions with this size.
>>
>> Please amend the date etc. at the top of the comment too.
>> Here are the results from non x86 worth adding I think:
> 
> Good point.  Done.
> 
>> POWER7 3.55GHz, revision 2.1 IBM,8231-E2B
>>    1024=1.3 GB/s
>>    2048=2.5 GB/s
>>    4096=4.8 GB/s
>>    8192=9.2 GB/s
>>   16384=16.8 GB/s
>>   32768=28.0 GB/s
>>   65536=41.4 GB/s
>>  131072=54.8 GB/s
>>  262144=40.0 GB/s
>>  524288=34.5 GB/s
>> 1048576=36.5 GB/s
> 
> Nice numbers. It'd be interesting to see power consumption :-)
> Can you determine RAM type and speed for that system?

dmidecode is not available on this linux ppc system,
though looking at specs for IBM Power 730 it says it takes
1066 MHz DDR3

The numbers above are over the throughput of that though,
so we're in cache land here anyway which each core on this system has:
L3(4MB), L2(256KB), L1d(32KB), L1i(32KB)

>> I have access to some NDA architectures and
>> there was an increase in performance seen there also.
> 
> I'd like to see a column or two for modern ARM-based server systems.

None I can quote unfortunately.

> Here's an updated patch. I'll update or remove the (RAM speed?)
> part before pushing.
> 

thanks,
Pádraig.