WLCG proposed a test period were all the VOs can exercise their computing models altogheter in a realistic scenario before the data taking. This exercise was named STEP09. The challenge ended on Friday 23:00 CET. A lot of activity happened during these last two weeks and the ATLAS Computing Model were fully exercised (finally!):

a) Huge load of MC production, exercising the farms and the data aggregation form T2s to the T1. The job type were basically G4 simulation jobs and AOD merging.

b) Huge load of Analysis jobs that were racing against the MC ones. I'd like to emphasize that this activity was quite well *disorganized* in overall... that was good as possibly could simulate quite well the possible impact of the uncontrolled user analysis that everyone will get very soon. There were general problems with Athena build jobs (compatibly libraries), the PanDA brokerage were bypassing the ATLAS release check at the sites (spotted last Thursday), the protocols to get data were basically lcg-cp for File stager and dcopen/file for remote connection... there's a big room for improvement in the Pilot Jobs based User Analysis, hopefully some tests will be performed in the coming months (use "natural SE" protocols for the FS, tune the Read Ahead Buffer -32Kb used during STEP-, etc.)

c) Non-Stop Reprocessing, this was the primary activity at the Tier-1s and was focused to perform multi-VO reprocessing at the same time fro those Tier-1 serving more than one experiment. That was achieved: ATLAS, CMS and LHCb managed to use the robots at the same time during several days. Six ATLAS sites manage to pass the metrics: reprocess at five times the data taking speed assuming 40% of LHC machine efficiency. And three sites manage to get gold stars, the metrics for this were: reprocess at five times data taking speed assuming 100% of LHC machine efficiency. PIC obtained gold star.

e) Data distribution: was broadly tested on top of the processing activity. Merged AOD were pre-placed at the sites: T1-T1-T2s, Functional Tests kept running all the time during the two weeks. There were no major incidences although some sites had some instabilities during the STEP: disk space emptied out, gridftp doors overloads (intensive LAN, WAN activity), but in overall data flow has shown its robustness and major improving with respect two years ago.

The STEP showed that manual work is still required but we managed to spot and learn which things would be the most demanding ones, the target now is to work in this direction and deploy an scalable operations model before the start.

Concerning our cloud, I'd say the exercise was a real success ! I want to thanks every party involved. Site has been stable (no major issues) that allowed STEP to pass through in a quite intensive way. I've barley seen an idle CPU all over the cloud, letting schedulers to battle against the different roles and jobs. Also the data transfer has been pretty high and the output form PIC to the Tier-2s has been quite stable at around 150-200MB/s.




MonteCarlo production ran between 1k and 1.5k jobs during STEP and efficiencies reached were well higher the 90%




On the other side User Analysis jobs were also filling the remaining slots, see the snapshot for the Pilot-based jobs (the ones directly submitted trough WMS are not accounted). Efficiencies ranked among the 70% and 80% depending on the sites playing the game, there were some issues (all understood!) that prevented pilot user analysis jobs to finish correctly.






PIC reprocessed the data two times, job efficiencies and data flow from the robot to the buffers and to the WN as a final step showed a very good performance. Pre-Stage is manage to keep up with 500 concurrent jobs which is beyond the required reprocessing speed for a Tier1 of small size. The interesting point is that reprocessing outputs were written to tape as well, this is what the computing model say, and we found no major problems with the simultaneous usage of drive for reading and writing.





What we learned during the STEP is that, once again, storage services are the most sensitive layer of the infrastructure. Small outage on the SE induce a general failure of every single activity. Storage services would have to be well dimensioned on those sites that had instabilities under constant load. We learned also that disk space is as volatile as it is in our laptops, a good prevision of the storage versus desired data is mandatory, it is true that ATLAS has to be more aggressive in data deletion but this is at the hands of the physics groups. I'd say that we should reinforce the Federative structure of our Tier-2s and take profit of these for the data distribution and always preserve some disk quantity in case of crisis. In the near future ATLAS will provide an interface that for modifying the shares at the sites so can be dynamic and not human-based, but also DDM will cross-check if the space token has enough free space before the data is shipped, hence preventing to replicate in case of shortages.

Once again my feeling and understanding is that our cloud is ready for data taking, let's hope data start to flow soon...

Tier-1 reliability in March and April

Now that everybody is talking about hot topics such as STEP09, let us take some break here to review some pretty outdated stuff. Sure we will be posting about the STEP09 in brief, but reviewing the scored monthly reliabilities and, most important, highlighting the causes of failures as homework to improve the service in the future has always proven to be a very useful exercise. Let's go then :-)
In March PIC scored 99% in reliability and 98% in availability. Pretty good, and above target. The missing reliability was caused by the jobs from a MAGIC user that filled up the local disk of WNs, turning them into black holes. Interesting lessons to learn here: protect ourselves from users filling up the disk (they will always do it, even if not deliberately) and minimise the impact that one user can have in all the other PIC users community.
April was not such a good month. Our reliability was above target (98%) but our availabilty was not (92%). The cause for the latter was our building yearly electrical maintenance. We scheduled two days downtime, and this brought us below the availability target for the Tier-1s. During this SD we tested a reduced downtime for the LHCb-DIRAC service. We managed to stop this service by just about 8 hours, so next time we should apply this to the other Tier-1 critical services.
The missing reliability in April was caused by a pretty bizarre problem in the SRM server. For some days we suffered huge overloads. The cause was finally found to be in the configuration of the dCache postgresql DB. In particular, the schedule of the "vacuum" procedures in the background. Using the "false" flag as recommended in the documentation was the origin of all our problems. After this incident, we have learnt quite a lot about postgres vacuum configuration and are pretty sure we are safe now. We have also learnt that trusting the documentation is not always a wise thing to do :-)

Important upgrades in last week SD: Improving the Computing Service

Last tuesday 19th May we had a Scheduled Downtime where quite a lot of important interventions were performed, aiming to improve the performance and reliability of some of the PIC services.
One of these interventions was the connection of the HP c7000 bladecenters to two stacked 10GE switches. Using a configuration already in place and originally designed for the dCache disk servers. The resulting bandwidth for the Computing LAN will be an average of 1,78 MB/s/core in the switch-router uplink and 3,9 MB/s/core in the bladecenter-switch uplink (after connecting each blacecenter with 4x1GE). One of the good things of this LAN infrastructure is its scalability, so we will keep an eye on the cacti monitoring of these links to anticipate wether we need to scale up.
Another important intervention which took place also affecting the Computing Service was the migration of the NFS shared software area to a new much more robust hardware: a FAS2020 cabin with SAS disks. This will not solve all the inherent problems that an NFS shared area brings to our lives, but at least will let us sleep a bit more relaxed while a more scalable solution for VO software access from the WNs arrives.

CPU ramp up

This week, on monday 4th May, the capacity of the Computing Service at PIC suffered a substantial increase. The number of available cores almost doubled in one go, so now we have a total of about 1400 cores. This corresponds to the deployment of 90 new (blade) servers, the MoU-2009 purchase of the Tier-1.
These new Worker Nodes have a L5420 Intel Xeon processor, which should give us better power consumption to specs ratio. This figure is important these days, when input power issues appear everywhere you go.
The first thing we wanted to check when powering on this new capacity was how stable was the temperature inside the machine room, and it looks that this has been ok. The other interesting issue is to see how well the Torque and Maui servers scale when doubling the number of nodes. We will need to keep an eye also in the scaling of the CEs...

February reliability, VOs joining the party

We should try and report about our scored availability and reliability for the past month of February... better before March is over!
Last month we got pretty good results for the OPS SAM tests: 98% availability (the missing 2% being basically the Scheduled Downtime on the 10th Feb) and... (drums in the background)... 100% reliability!
Yes. There we are, at the top of the ranking. Well, actually CERN, FNAL, BNL and RAL did also score max reliability last month, together with us.
Somebody might argue (the experiments, actually) that this are OPS VO tests, so not reflecting reality. Actually, the opposite was true since not long ago. The experiments SAM tests were still being put together, and they were showing lots of false negatives.
Anyway, the results for the VO-specific SAM tests in February were also pretty good at PIC: 98% reliability for ATLAS, 99% for CMS and 96% for LHCb. The ATLAS and CMS detected unreliability is completely true, we must admit. This was an indicent we suffered on the 22nd February (Sunday, by the way) with our "sword of damocles", also known as the SGM-dedicated WN. Some jobs hanged there, completely blocking further execution of the SGM jobs from the VOs. Luckily, Arnau reads his e-mail during weekends, and the problem was detected and solved quite fast (thanks!).
The LHCb reliability number is not really reliable, since their SAM test framework had some hickups the first 5 days of the month.
We had recently the WLCG Workshop in Prague, there we heard big complaints from CMS saying the reliability of Tier-1s is not good at all. The numbers they showed were indeed quite bad, and this is mostly due to the fact that they are adding extra ingredients to their reliability calculation. In particular, they use the results of routine dummy job submissions (JobRobot), and it happened that, even if our CMS SAMs were strong green, the JobRobot was more reddish.
I think it is good that experiments make their sites monitoring more sophisticated. However, for this to be a useful tool in improving reliability... they first need to tell us!
Now we are

CMS cpu consumer, back to business

Looks like someone in CMS was reading this blog, since few hours after the last post, saturday evening, CMS jobs started arriving to PIC. We see a constant load of about 300 jobs from CMS since then. Not bad.
Apparently these are the so-called "backfill" jobs. All the Tier-1s but us (and Taiwan, down these days due to a serious fire incident) started running these backfill jobs early March. After a bit of asking around, we found out that PIC was not getting its workload share because the old 32bit batch queue names were hardcoded somewhere in the CMS sytem (we deprecated 32bit queues more than one month ago!) plus they had a bug in the setup script that got the available TMPDIR space wrong.
Good that we found these problems and that they were promptly solved. Now CMS is back to the cpuburning business at PIC. ATLAS is still debugging the memory-exploding problem that stopped jobs being sent to PIC about one week ago. Looks we are close to the solution (missing packages) and we will soon se both experiments competing again for the CPU cycles at PIC.

CPU delivery efficiency

We are these days collecting the accounting data for February 2009. Looks like we reached a record figure for CPU efficiency delivery last month. The 3 LHC experiments used up to 80% of the total CPU days available at PIC: almost 37.000 ksi2k·days. This was largely thanks to ATLAS, who consumed around 80% of those CPU days. LHCb used just above 15% and CMS a mere 5%.
So, well done for ATLAS. It is true that most of that load are not "Tier-1 type jobs", but just contribution to the experiment MonteCarlo production. Anyway, it is better that Tier-1 resources are used for simualation rather than stay idle consuming electricity, heating the computing room and watching their 3-year lifetime pass by (at a rate of about 6 kEur/month).
From our point of view the Panda system which is used in ATLAS, and that implements the now so-loved pull model for computing (or pilot jobs), is definetely doing a good job in consuming all available CPU resources.
Unfortunately, not everything is so nice. This last week we have seen the CPU utilisation at PIC decreasing quite a lot. The ATLAS Panda system was not sending jobs to PIC, and we discovered this was due to a problem with Athena software running in 64bit OS. Suddenly the production jobs running at PIC's SL4/64bit WNs exploded in memory utilisation and were eventually killed by the system. The experts are working now to understand and fix this, hope they find a patch soon.
Let's see when CMS and LHCb implement efficient CPU consuming systems similar to ATLAS and we can benefit of being a multi-experiment Tier-1.
Meanwhile, at PIC, idle CPUs are transforming electricity into heat. Waiting for ATLAS to cure their 64bit indigestion.