FCPPL computing meeting

lundi 12 déc. 2011, 10:30 → 12:30 Europe/Paris

Meeting to follow up projects initiated between IHEP Computing Center and CC-IN2P3

10:30 → 11:30 Network Performances and Monitoring

slides 2011-12-12-FCPPL-Trend_Data_Transfer.pdf

Throughput between CC and IHEP

The throughput for the network is worse in the direction IHEP-to-CC than for CC-to-IHEP.

There seems to be however related with a more general problem: Jérôme is investigating with RENATER for US and japanese Tier2s. There could be something wrong with the GEANT backbone. Japanes have observed different throughput with European Tier1s. Transfers with France and UK are worse than with CERN, Spain or Italy. Investigation still going on.

Perfsonar Setup and Monitoring

Perfsonar is setup on both sides at IHEP and at CC. there are tests every 4 hours for 60 seconds. However, tests between IHEP and CC are inactive since november the 28th as can be seen on the CC- personar graphs page:
https://ccperfsonar-lhcopn.in2p3.fr/serviceTest/index.cgi?eventType=bwctl
Is it a firewall problem or something else?
For information, the Tier1 perfsonars dashboard is under: https://130.199.185.78:8443/exda/?page=25&cloudName=LHCOPN

Network data analysis

Shan has setup the tools to extract the collected data by the router and store them in text format. Fabio and Shan are exploring 2 ways of storing them for high-throuhput analysis: one is using a Hadoop cluster and the other is using Redis, an in-memory key-value store. Currently they are developing the necessary tools to make some tests using real sample data in order to draw conclusions regarding the best way to store and summarize the collected data. The problem of graphical visualization of the summarized data has not been addressed yet.


David is providing the FTS log files through an Apache server every 15 minutes. The data is stored for 11 consecutive days. The logs can be found in http://cclcgftmli02.in2p3.fr/transferlogs
The idea is to cross check the activity at the level of the router with the activity shown in the logs of FTS.

Actions

- continue investigation on degraded performance of the network between CC and remote sites.
- Why no more tests of perfsonar since 28th of November?
- Continue setting up of the analysis of the routers data and later cross checks with FTS data.

10:30 Status of Perfsonar installation and monitoring

10:50 Real Data exchange Monitoring and FTS information sharing

11:30 → 12:30 VO France-Asia

VO France-Asia deployment and usage It is deployed at CC, KEK and KISTI. At Kisti, a new PhD student has deployed the Dirac service. The Dirac server is in Hanoï. This initiated a collaboration also with Vietnam. André Tilquin, affiliated to CPPM and THCA (Tsinghua Center for Astrophysics, http://www.thca.tsinghua.edu.cn/en/index.php/Main_Page), has some computing needs that seem well suited for the intended use-cases of the France-Asia VO. He has already experience running his tasks on the grid under the umbrella of the ESR (Earth Science Research) VO, so his software is grid-ready. Fabio met him in Beijing and suggested him to join the France-Asia VO in order to perform feasibility tests. His jobs are CPU-bound and he needs to regularly submit campaigns of several hundreds of them. Fabio addressed André to Yonny who agreed to take care of setting things up and assisting André performing the foreseen tests. A question was raised on the plans of IHEP to join the VO. Yonny explains that this needs small set of machines and little storage to begin with. For example, there are actually about 3000 cores on the VO of which 80% are from CC, the rest from KEK and KISTI. VO Web Site The VO web site is not yet ready. Yonny will work with webmaster at CC to set it up. VO Tests   As shown in the attached slides, a suspiciously low performance has been observed when transferring files from IHEP to CC-IN2P3. Memory-to-memory network tests based on iperf showed that there is currently a severe limitation: the observed throughput is less than 1 MB/sec which is considered very low taking into account that the link from China to Europe is 2 x 1 Gbps.   The cause of this problem is not yet understood but could be related to the problem highlighted in the previous discussion.   Once this issue is solved, the file transfer throughput tests will be resumed in order to identify the ability of iRods to exploit the available bandwidth. In the particular case of TREND, there are too many small files which does not help to reach good throughput in long distance transfers.  There was a suggestion to use tarred large file as this is possible with IROD's and to move to UDP protocol for transfers.   Regarding the usage of France-Asia VO by TREND, Fabio intends to perform some initial tests of job submission and data access to the iRods repository at CC-IN2P3 in January 2012.   Actions: -Build the VO site. -Check for IHEP the possibility to join the VO with some resources. -check the source of low transfers: disentangle between network and Irod's effects. - Send test jobs to access the data. -Setup the new user (André Tilquin) to run on the VO

11:30 VO France-Asia Status of deployement