[Gluster-users] Remove an artificial limitation of disperse volume

[Xavier Hernandez]

On 07/02/17 17:29, Olivier Lambert wrote:
> Yep, but if I hit a 30% penalty, I don't want that :) Any idea of the
> perf impact? I'll probably contact Xavier directly if he's not here!

It depends on the workload. The basic problem is that the I/O write size 
must be a multiple of the number of data bricks to avoid 
read-modify-write cycles (i.e. they write always full stripes, so no 
need to read anything from disk).

Normally applications issue writes that are a power of 2 (512, 4096, 
131072, ...), so the best configurations for ec should also be powers of 
two. But of course you can have bad performance with ec even if you use 
a power of two if the application ends writing blocks of 1000 bytes for 
example.

Anyway, I cannot give you real numbers. Long time ago I did some tests 
but I was unable to get conclusive results. Probably I was having too 
many interferences from other translators that were hiding the real 
performance impact (maybe write-behind). This is good because this means 
that peak performance could be the roughly the same with "optimal" and 
"non-optimal" configurations, but sustained throughput should be worse 
for "non-optimal" configurations because at some point the caching made 
by write-behind and other xlators will become full and they won't be 
able to continue hiding the performance hit.

Some day I'll try to find time to get more accurate results regarding 
this issue, but for now only thing I can say is to test your specific 
workload with your specific configuration and see if it suffers a 
performance impact or not.

Xavi

>
> On Tue, Feb 7, 2017 at 5:27 PM, Jeff Darcy <jdarcy at redhat.com> wrote:
>>
>>
>> ----- Original Message -----
>>> Okay so the 4 nodes thing is a kind of exception? What about 8 nodes
>>> with redundancy 4?
>>>
>>> I made a table to recap possible configurations, can you take a quick
>>> look and tell me if it's OK?
>>>
>>> Here: https://gist.github.com/olivierlambert/8d530ac11b10dd8aac95749681f19d2c
>>
>> As I understand it, the "power of two" thing is only about maximum
>> efficiency, and other values can work without wasting space (they'll
>> just be a bit slower).  So, for example, with 12 disks you would be
>> able to do 10+2 and get 83% space efficiency.  Xavier's the expert,
>> though, so it's probably best to let him clarify.
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org
> http://lists.gluster.org/mailman/listinfo/gluster-users
>