Share
Embed
Grid Monitoring at CERN with Elastic
In this talk Pablo Saiz presents the five different use cases where Elastic is used for WLCG monitoring at CERN:
- Messaging
- Job Monitoring
- Data Monitoring
- Infrastructure Monitoring
- Cloud Benchmarking
Pablo will also highlight the future goals with Elastic and their plan to expand their Elastic useage at CERN.
Pablo Saiz | Elastic{ON}Tour Munich | November 10, 2015
Elastic Co
November 10, 2015
Transcript
29 slides
-
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN Table of
contents ▪ CERN ▪ Worldwide LHC Computing GRID (WLCG) ▪ Elastic for WLCG Monitoring ▪ Messaging ▪ Job Monitoring ▪ Data Monitoring ▪ Infrastructure monitoring ▪ Cloud benchmarking 05/11/15 2 -
IT-SDC 3 CERN was founded 1954: 12 European States “Science
for Peace” Today: 21 Member States Member States: Austria, Belgium, Bulgaria, the Czech Republic, Denmark, Finland, France, Germany, Greece, Hungary, Israel, Italy, the Netherlands, Norway, Poland, Portugal, Slovakia, Spain, Sweden, Switzerland and the United Kingdom Candidate for Accession: Romania Associate Member in Pre-Stage to Membership: Serbia Applicant States for Membership or Associate Membership: Brazil, Cyprus, Pakistan, Russia, Slovenia, Turkey, Ukraine Observers to Council: India, Japan, Russia, Turkey, United States of America; European Commission and UNESCO ~ 2300 staff ~ 1600 other paid personnel ~ 10500 users Budget (2014) ~1000 MCHF Well known for: • Physics • WWW • Media: Angels & Demons, Flashforward, Daily show -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN CERN’s Accelerator
Complex 05/11/15 4 http://cernland.net -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN Colliding particles
(event) 05/11/15 5 Black Hole Higgs -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN WLCG ▪
Biggest scientific Grid project in the world ▪ ~170 computer centers (site) ▪ 1 Tier 0 (distributed in two locations) ▪ 12 bigger centers (Tier 1) ▪ ~160 Tier 2 ▪ 42 countries ▪ 10,000 users ▪ Running since Oct 2008 ▪ 2 million jobs per day ▪ ~600.000 cores ▪ 300 PB data 05/11/15 6 http://cern.ch/wlcg -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN WLCG 05/11/15
7 Including Max-Planck- Institut für Physik , Munich -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN WLCG Monitoring
▪ Many different tools: ▪ Job, data transfers, services ▪ Usual Architecture: ▪ Python, RDBMS, apache, javascript ▪ Running since 2007 ▪ Team of ~10 people ▪ Different audience: ▪ Site/service/experiment managers, end users, general public 09/11/15 8 http://dashboard.cern.ch -
IT-SDC Elastic{ON} Tour in Munich, Pablo Saiz, CERN ElasticSearch configuration
▪ 2 clusters ▪ Production: 8 data, 2 search, 3 master ▪ Development: 2 data, 1 search, 1 master ▪ Data nodes: physical, 32GB, 32 core ▪ Search, master: virtual 8GB ▪ ~10 different use cases -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN Use case
1: Messaging Team (MIG) ▪ Goal: ▪ Check the status of the machines and services used for Messaging ▪ 25 machines, 15 clusters, ~130 applications, 260 M msg/day ▪ Using ElasticSearch and Kibana 3 in production since Aug 2014 ▪ 3 daily indexes, ~3M documents (850MB) per day ▪ Documents contain sensitive information ▪ Ensure only authorized people can access them 05/11/15 10 Messaging infrastructure ElasticSearch ESPER Transport layer (Messaging) Log Check Metric Status Graphite http://cern.ch/messaging -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN Use case
2: Job Monitoring ▪ Goal: follow up the status of job processing ▪ ~2M daily jobs (170M updates) over ~170 sites ▪ Multiple applications: interactive view, summaries, accounting… ▪ Landing page query with 2 million records aggregation ▪ Aggregation 60M entries (100GB) ▪ In production (with RDBMS) since 2008 ▪ Elasticsearch alternative in place ▪ Currently checking data consistency 05/11/15 12 -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN Accessing raw
data / summary 05/11/15 13 -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN Architecture ▪
Python collectors ▪ Apache web server ▪ With mod_python ▪ Javascript ▪ Datatables, highcarts… ▪ Currently replacing the storage with ES 05/11/15 14 -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN ElasticSearch result
▪ Good speed-up in job parsing: ▪ From 220 jobs per second to 580 (2,6x) ▪ Better speed-up in applications: ▪ From 12 seconds to 0,70s in the most used query ▪ More queries possibilities: ▪ More than one month of aggregation 05/11/15 15 -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN Performance 05/11/15
16 Opening the landing page lasticSearch S (no cache) 0 3 6 9 12 11.8 0.6 Maximum updates per second Python +ElasticSearch Python +RDBMS 0 150 300 450 600 223 584 One Month Query lasticSearch S (no cache) 0 300 600 900 1200 1084.3 14.1 -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN Use case
3: Infrastructure Monitoring ▪ Goal: ▪ Follow up status of sites/services using user defined metrics ▪ 2M entries per day (1M raw metrics, 1 M combined metrics) ▪ Metrics: ▪ Numerical or string ▪ Non-linear in time (ndependent sources, future value, changing old entries) ▪ Keeping only status changes ▪ Combining metrics (AND/OR/ANY/ALL/FILTER) ▪ Creating ‘Views’ as lists of metrics ▪ Use in production (RDBMS) since 2010 ▪ ES alternative ready for data insertion and combining metrics ▪ Moving from python to java ▪ Using JEST interface ▪ Using ESPER ▪ ESPER/ES reduce significantly latency of metric combination ▪ UI needs to be modified 05/11/15 17 http://wlcg-mon.cern.ch/ -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN How to
combine metrics 09/11/15 19 Submit: ce1, ce2 OK ce3: FAIL Topo: Site1: ce1, ce2, St1, Tr1 Sote2: ce3, St2 Downtime ce3: maint CE ce1: OK ce2, ce3:FAIL Storage St1: OK St2: FAIL Transfer Tr1:OK CE Site1: Ok Site2: MAINT Storage Site1: OK Site2: FAIL Transfer Site1: OK Storage+ Transfer Site1:OK Site2: FAIL Status Site1: OK Site2: FAIL HOST SITE Example: Any CE AND ( Storage OR Transfer) Where CE=Submit AND Efficiency>80; Storage=Read AND Write; Transfer=CanTransfer Downtimes; Topology Efficiency: ce1:85 ce2:60 ce3:90 Read: St1: Ok St2: OK Write: St1 OK St2: FAIL Transfer: Tr1 OK CE ce1: OK Ce2: FAIL Ce3:MAINT Storage St1: OK St2: FAIL Transfer Tr1:OK -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN Use case
4: Data Monitoring ▪ Goal: ▪ Provide speed, success rate, volume of data movements ▪ 25M daily transfers, avg speed 10 GB/s ▪ RDBMS version in production since 2010 ▪ Elastic alternative ▪ 1 index: 160M documents, 80 GB ▪ Current status: ES index created, and being populated. Still to work on UI 05/11/15 20 -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN WLCG Transfers
dashboard 05/11/15 21 -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN Use case
5: Benchmarking Commercial Cloud Resources ▪ Goal: Compare different benchmark tools on different cloud providers ▪ Use 6 different benchmark tools and X cloud providers ▪ 1 index: 270K documents (~250 MB) ▪ Detailed analysis performed with Ipython analytics http://indico.cern.ch/event/384358/session/12/contribution/25 05/11/15 22 -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN Benchmark results
at a glance ▪ Metric ~ [s]; each point ⬄10 min average; Colour ⬄ Cloud 05/11/15 23 -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN Qualitative look
at data 05/11/15 24 1 VM 16 VMs 30 VMs 16 VMs 25 VMs 20 VMs ▪ Identifiable transition of CPU performance when load changes ▪ Seen in all benchmark measurements. Performance recovers scaling down -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN Qualitative look
at data ▪ Larger dispersion in KV and FastBmk values in the highest- load region 05/11/15 25 1 VM 16 VMs 30 VMs 16 VMs 25 VMs 20 VMs -
IT-SDC IT-SDC Analysis done outside elastic ▪ Correlation between different
fields in documents ▪ Histograms with floating numbers 05/11/15 26 Evolution of a single VM in the parameter space FastBmk Vs KV Ratio mean(30VMs)/mean(16 VMs) Projection-X Aggr. x hypervisor Projection-Y Aggr. x hypervisor Profile-X 2D plot KV FastBmk A single Hyperv.!! -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN Next goals
▪ Use ES for Job Monitoring in production ▪ Migrate to ES 2 ▪ Check indexes/shards ▪ Ensure fine-grained security ▪ Currently, based on apache configuration (allow only REST interface!) ▪ Need to move to another alternative ▪ Move to Kibana 4 (once security in place) 05/11/15 27 -
IT-SDC IT-SDC Elastic{ON}Tour in Munich, Pablo Saiz, CERN Summary ▪
WLCG Grid Monitoring at CERN has been using Elastic in production for over year ▪ Currently, evaluating ElasticSearch for more use cases ▪ Very promising results so far! ▪ Looking forwards to expanding our Elastic usage! 05/11/15 28