Fabric Monitor, Accounting, Storage and Reports experience at the INFN Tier1 Felice Rosso on behalf of INFN Tier1 Workshop sul calcolo e reti INFN - Otranto

Outline CNAF-INFN Tier1 FARM and GRID Monitoring Local Queues Monitoring –Local and GRID accounting Storage Monitoring and accounting Summary

Introduction Location: INFN-CNAF, Bologna (Italy) –one of the main nodes of GARR network Computing facility for INFN HNEP community –Partecipating to LCG, EGEE, INFNGRID projects Multi-Experiments TIER1 –LHC experiments (Alice, Atlas, CMS, LHCb) –CDF, BABAR –VIRGO, MAGIC, ARGO, Bio, TheoPhys, Pamela... Resources assigned to experiments on a yearly Plan.

The Farm in a Nutshell - SLC 3.0.6, LCG 2.7, LSF ~ 720 WNs LSF pool (~1580 KSI2K) -Common LSF pool: 1 job per logical cpu (slot) -MAX 1 process running at the same time per job –GRID and local submission allowed On the same WN can run GRID and not GRID jobs On the same queue can be submitted GRID and not GRID jobs –For each VO/EXP one or more queues –Since 24th of April jobs were executed on our LSF pool (~ GRID) –3 CEs (main CE 4 opteron dualcore, 24 GB RAM) + 1 CE gLite

Access to Batch system “Legacy” non-Grid Access CELSF Wn1WNn SE Grid Access LSF client UI Grid

Farm Monitoring Goals Scalability to Tier1 full size Many parameters for each WN/server DataBase and Plots on Web Pages Data Analysis Report problems on Web Page(s) Share data with GRID tools RedEye: INFN-T1 tool monitoring RedEye: simple local user. No root!

Tier1 Fabric Monitoring What do we get? CPU load, status and jiffies Ethernet I/O, (MRTG by net-boys) Temperatures, RPM fans (IPMI) Total and type of active TCP connections Processes created, running, zombie etc RAM and SWAP memory Users logged in SLC3 and SLC4 compatible

Tier1 Fabric Monitor

Local WN Monitoring On each WN every 5 min (local crontab) infos are saved locally ( 2-3 TCP packets) 1 minute later a collector “gets” via socket the infos –“gets”: tidy parallel fork with timeout control To get and save locally datas from 750 WN ~ 6 sec best case. 20 sec worst case (timeout knife) Upgrade DataBase (last day, week, month) For each WN --> 1 file (possibility of cumulative plots) Analysis monitor datas Local thumbnail cache creation (web clickable)

Web Snapshot CPU-RAM

Web Snapshot TCP connections

Web Snapshot users logged

Analyzer.html

Fabric  GRID Monitoring Effort on exporting relevant fabric metrics to the Grid level e.g.: –# of active WNs, –# of free slots, –etc… GridICE integration –Configuration based on Quattor Avoid duplication of sensors on farm

Local Queues Monitoring Every 5 minutes on batch manager is saved queues status (snapshot) A collector gets the infos and upgrades the local database (same logic of farm monitoring) –Daily / Weekly / Monthly / Yearly DB –DB: Total and single queues 3 classes of users for each queue Plots generator: Gnuplot 4.0

Web Snapshot LSF Status

UGRID: general GRID user (lhcb001, lhcb030…) SGM: Software GRID Manager (lhcbsgm) OTHER: local user

UGRID: general GRID user (babar001, babar030…)SGM: Software GRID Manager (babarsgm)OTHER: local user

RedEye - LSF Monitoring Real time slot usage Fast, few CPU power needed, stable, works on WAN RedEye simple user, not root BUT… 1.all slots have the same weight (Future: Jeep solution) 2.Jobs shorter than 5 minutes can be lost SO: We need something good for ALL jobs. We need to know who and how uses our FARM. Solution: Offline parsing LSF log files one time per day (Jeep integration)

Job-related metrics From LSF log file we got the following non-GRID info: LSF JobID, local UID owner of the JOB “any kind of time” (submission, WCT etc) Max RSS and Virtual Memory usage From which computer (hostname) the job was submitted (GRID CE/locally) Where the job was executed (WN hostname) We complete this set with KSI2K & GRID infos (Jeep) DGAS interface

Queues accounting report

KSI2K [WCT] May 2006, All jobs

Queues accounting report CPUTime [hours] May 2006, GRID jobs

How we use KspecINT2K? -1 slot → 1 job - -For each job:

KSI2K T1-INFN Story

Job Check and Report Lsb.acct had a big bug! –Randomly: CPU-user-time = 0.00 sec –From bjobs -l correct CPUtime –Fixed by Platform at 25th of July 2005 CPUtime > WCT? --> Possible Spawn RAM memory: is a job on the right WN? Is the WorkerNode a “black hole”? We have a daily report (Web page)

Fabric and GRID monitoring Effort on exporting relevant queue and job metrics to the Grid level. –Integration with GridICE –Integration with DGAS (done!) –Grid (VO) level view of resource usage Integration of local job information with Grid related metrics. E.g.: –DN of the user proxy –VOMS extensions to user proxy –Grid Job ID

GRID ICE Dissemination GridICE server (development with upcoming features) GridICE server for EGEE Grid GridICE server for INFN-Grid

GRID ICE For each site check GRID services (RB, BDII, CE, SE…) Check service--> Does PID exist? Summary and/or notification From GRID servers: Summary CPU and Storage resources available per site and/or per VO Storage available on SE per VO from BDII Downtimes

GRID ICE Grid Ice as fabric monitor for “small” sites Based on LeMon (server and sensors) Parsing of LeMon flatfiles logs Plots based on RRD Tools Legnaro: ~70 WorkerNodes

GridICE screenshots

Jeep General Purpose collector datas (push technology) DB-WNINFO: Historical hardware DB (MySQL on HLR node). KSI2K used by each single job (DGAS) Job Monitoring (Check RAM usage in real time, efficiency history) FS-INFO: Enough available space on volumes? AutoFS: all dynamic mount-points are working? Match making UID/GID --> VO

The Storage in a Nutshell Different hardware (NAS, SAN, Tapes) –More than 300 TB HD, 130 TB Tape Different access methods (NFS/RFIO/Xrootd/gridftp) Volumes FileSystem: EXT3, XFS and GPFS Volumes bigger than 2 TBytes: RAID 50 (EXT3/XFS). Direct (GPFS) Tape access: CASTOR (50 TB of HD as stage) Volumes management via Postgresql DB 60 servers to export FileSystems to WNs

Storage at T1-INFN Hierarchical Nagios servers to check services status –gridftp, srm, rfio, castor, ssh Local tool to sum space used by VOs RRD to plot (volume space total and used) Binary and owner (IBM/STEK) software to check some hardware status. Very very very difficult to interface owner software to T1 framework For now: only report for bad blocks, disks failure and FileSystem failure Plots: intranet & on demand by VO

Tape/Storage usage report

Summary Fabric level monitoring with smart report is needed to ease management T1 has already solution for 2 next years! Not exportable due to man-power (no support) Future at INFN? What is T2s man-power? LeMon&Oracle? What is T2s man-power? RedEye? What is T2s man-power? Real collaboration is far from mailing list and phone conferences only