It just occurred to me that even though you didn't change the BDII on
purpose you changed the system moving from CREAM to ARC as well as
changing the WNs.
The numbers in REBUS started to fluctuate wildly in May, before they
were stable on 43243/2912 = 14.85. I've asked REBUS people to check, but
that might explain some of the differences.
cheeers
alessandra
On 17/11/2016 14:33, Alessandra Forti wrote:
> Hi Marcus,
>
> thanks for confirming. It is still not clear to me why REBUS sees this
> wild variations for ECDF. I'll try to get an answer from them.
>
> cheers
> alessandra
>
>
> On 17/11/2016 14:20, Marcus Ebert wrote:
>> Hi Alessandra,
>>
>>
>> On Thu, 17 Nov 2016, Alessandra Forti wrote:
>>
>>> size of the site is manually inserted in the BDII. I agree in ECDF
>>> it is variable but you really should put a meaningful value that
>>> averages to a meaningful HS06 number. I thought you did that but
>>> ECDF is red again. This time APEL is bigger than ATLAS. You seem to
>>> change the capacity in the BDII every month [1] can you confirm
>>> that? You should put values whose ratio is ~HS06 you publish.
>>>
>> No, I don't think it was changed every month. It was changed in
>> October to make it consistent between the 2 numbers we report and to
>> reflect the current worker node systems we run on (ringfenced nodes,
>> general ECDF cluster, Openstack - all with different HepSpec and job
>> slots/cores).
>> (see below)
>> This value should reflect the different systems we are running on in
>> very good approximation now.
>>
>>> aforti@vm7>site=UKI-SCOTGRID-ECDF; ldapsearch -LLL -x -h
>>> top-bdii.tier2.hep.manchester.ac.uk:2170 -b
>>> "mds-vo-name=${site},mds-vo-name=local,o=grid" | perl -p00e 's/\r?\n
>>> //g'|egrep -i 'bench|spec|logical'
>>> GlueHostBenchmarkSF00: 0
>>> GlueHostBenchmarkSI00: 0
>>> objectClass: GlueHostBenchmark
>>> GlueHostProcessorOtherDescription: Cores=8, Benchmark=12.9-HEP-SPEC06
>>> GlueSubClusterLogicalCPUs: 528
>>>
>> That's the updated correct one. It was updated in October, so I think
>> we should wait for the November numbers once the whole month is over.
>> Cores and Hepspec are averaged over the different systems taking the
>> different number of cores/machines into account we really run on.
>>
>>
>>> ATM REBUS reports weird stuff not corresposnding to 12.9
>>>
>>> October: 111945/9570 =11. 69 <-- atlas claims 11.884 until August
>>> included
>>> September: 74195/7040 = 10.54 <-- atlas see 10.5 from September
>>> onward in line with this numbers
>>> October: 76167/7291=10.44 <-- similar enough
>>> November: 6811/528 = 12.89 <-- this is ok if ATLAS sees it, but I
>>> suspect numbers are not updated that often and it might be a
>>> discrepancy again.
>>>
>> Atlas sees 10.5 because that's what we my mistake reported. We didn't
>> updated the Glue value and only the one for APEL when we added new
>> worker nodes. 10.5 was the wrong, too low value. Since we updated now
>> the APEL and GLUE value to be consistent, there should be no
>> reason/possibility that ATLAS sees something different for November.
>>
>>> so there are 3 points here
>>>
>>> 1) Do you update your numbers to maintain the HS06 ratio in the BDII
>>> consistently? I don't think changing numbers monthly is a good idea
>>> but they should at least match the HS06 value.
>> No, we don't change monthly.
>> We only looked into it because of the discrepancy you reported and
>> found that a) that the 2 different values we report, Apel and Glue
>> one, are not consisten with each other, b) both don't reflect the new
>> hardware we are running on since a while for the SL6 analysis queue.
>> That's why it was changed in October. Before I think the last change
>> was in July when we got new machines to run on (differently
>> configured for job slots than our ringfenced nodes which made a
>> change neccessary)
>> The change in October reflected the addition of the Openstack nodes
>> for the SL6 queue.
>>
>>> 2) If you do that why rebus is reporting a different set of numbers
>>> for example I'd expect Ocotber 7291*12.9 = 94053 not 76167
>> We don't do that.
>> It was changed in October, so probably that's why it's different
>> since it was not the same for the whole month?
>> I would expect that November onwards it should now correspond to 12.9
>>
>>> 3) ATLAS doesn't seem to update the HS06 often enough to have such
>>> frequent changes. And TBF most sites usually don't change their size
>>> every month.
>>>
>> As I said, we also don't do that.
>>
>>
>> I think we should wait until the end of October to see if it will be
>> green then and consistent.
>> In any case, we will look through the published data using the
>> scripts you published to make sure it will be consistent in the future.
>>
>>
>> Cheers,
>> Marcus
>>
>>> [2] http://tinyurl.com/j2fylyx
>>>
>>> On 17/11/2016 12:05, Marcus Ebert wrote:
>>>> Thanks Alessandra,
>>>>
>>>> I think I understand now, also from previous discussions in the
>>>> list here.
>>>> Basically, it only tests if 2 values published by a site, both
>>>> defined in
>>>> the bdii and put in manually by the site, agree or not, but
>>>> doesn't say
>>>> anything about the correctness of the HEPSPEC value used.
>>>> So it seems what really meaningfully can be compared is just the
>>>> wallclock
>>>> work from Atlas and APEL, if it's not scaled at a site.
>>>>
>>>> Wouldn't it be better then to split the plot in 2 different ones,
>>>> - one for the ratio of wallclock hours Atlas/APEL to have a site
>>>> check
>>>> that both values published are consistent, and
>>>> - second one only for the wallclock work ratio Atlas/APEL to see any
>>>> differences between the reported wallclock work in APEL and the ATLAS
>>>> records?
>>>>
>>>> If it shows for example "red" right now, it's not obvious just
>>>> from the
>>>> plot which of the 2 numbers are the problem.
>>>>
>>>>
>>>> Cheers,
>>>> Marcus
>>>>
>>>> On Tue, 8 Nov 2016, Alessandra Forti wrote:
>>>>
>>>> > Hi Marcus,
>>>> > > > Thanks, I think I nearly understand it now. To fully
>>>> understand, > could you please explain how HS06 in Atlas wallclock
>>>> work is determined? > It isn't the same that > is used in APEL
>>>> wallclock work, is it?
>>>> > > the presentation I gave yesterday at the HEPSYSMAN gives the
>>>> details
>>>> > >
>>>> https://indico.cern.ch/event/577279/contributions/2353919/attachments/1367099/2071452/20161107_hepsysman-accounting.pdf
>>>> > > > in the specific today I've also started an FAQ
>>>> > >
>>>> https://twiki.cern.ch/twiki/bin/view/LCG/AccountingFAQ#How_are_the_ATLAS_numbers_in_SSB
>>>> > > > cheers
>>>> > alessandra
>>>> > > On 01/11/2016 09:52, Marcus Ebert wrote:
>>>> > > Hi Alessandra,
>>>> > > > > On Tue, 1 Nov 2016, Alessandra Forti wrote:
>>>> > > > > > > I'm not sure if I understand it or if it makes sense
>>>> that way:
>>>> > > > > Basically what you are saying is that the initial number
>>>> values
>>>> > > > > "HS06 on the atlas dashboard, HS06 in APEL, ratio,
>>>> wallclock > > in > > ATLAS,
>>>> > > > > wallclock in APEL, wallclock ratio"
>>>> > > > > are really
>>>> > > > > "wallclock work in the Atlas, wallclock work in APEL,
>>>> ratio, > > > > wallclock
>>>> > > > > work in Atlas (unscaled), wallclock work in APEL (maybe
>>>> > > > > scaled)",
>>>> > > > > isn't it?
>>>> > > > the fields are
>>>> > > > > ATLAS wallclock work (HS06*hours), APEL wallclock work >
>>>> > (HS06*hours), > ratio, ATLAS wallclock (hours), APEL wallclock
>>>> (hours > > maybe internally > scale), ratio
>>>> > > > > > Thanks, I think I nearly understand it now. To fully
>>>> understand, could > > you
>>>> > > please explain how HS06 in Atlas wallclock work is determined?
>>>> It > > isn't
>>>> > > the same that is used in APEL wallclock work, is it?
>>>> > > > > > > Cheers,
>>>> > > Marcus
>>>> > > > > > > > >
>>>
>>>
>>
>
--
Respect is a rational process. \\//
Fatti non foste a viver come bruti (Dante)
|