First Edition: 6 Dec 2019

Version 1.1 (10:50 PT 6 Dec 2019)

Power outage time-line.  

[Tentative proposal]  Not many details are currently known, but this power outage will affect substations #7 (next to bldg 50) and #8 (located on the 4th floor of bldg 50).  All of bldg 50 will be without normal power.  The facilities (F&O) group plan to do their maintenance during the 4-day period starting 26 Dec 2019.  However, the outage will start earlier due to lack of staff during the holiday shutdown.  Minimally, it is expected that all H.A. (High Availability) and experiment-critical equipment will be powered throughout the 16+ days of the holiday shutdown.  This page captures what Fermi will need to maintain a minimal data processing effort running during the outage.

Outage preparation task list

• Define needed xrootd resources (Wilko Kroeger)

• Confirm sufficient xrootd space to handle 16+ day HPSS outage (Wilko Kroeger)

• Define needed Level 0 and half-pipe resources (Steve Tether)
• Define needed Level 1 resources (Tom Stephens)
• Update service → VM → hypervisor host mapping (Brian Van Klaveren)
• Suggest/Move non-HA VMs to HA (Brian Van Klaveren)
• Define needed ASP resources (Jim Chiang)
• Define needed resources for critical Fermi Science pipelines (various)
  • Gravitational wave analysis (Nicola Omodei)
  • Flare Advocates (Gulli Johannesson, Stefano Ciprini)
  • Burst Analysis (Dan Kocevski)
     

LISOC Operations Functions

Table of LISOC Tasks and Services

The following table of servers must remain powered up and operational for Fermi Level 1 and critical Science Pipelines to function.

Fermi has requested that all VMs be relocated (at least temporarily) to the two H.A. hypervisor machines, thus some of the tasks listed below are no longer relevant.

• Confirm current H.A. rack occupants.  spreadsheet from Christian Pama
  Old (2017) spreadsheet here
• Confirm the VM-master for a given VM.  Use the 'node' command, e.g., $ node -whereis fermilnx-v12 (obsolete)
• Confirm the tomcat <-> service associations.  Table here.
• Confirm the tomcat-VM associations in this table. Use the 'node' command, e.g., $ node -whereis glast-tomcat01
  
  

† Equipment categories

Total non-HA machines to receive emergency power:

 Note that HPSS is NOT required by Fermi.

The services for L1:

oracle

• pipeline
• data catalog
• group manager

mysql

• calibrations

tomcats

• pipeline
• data catalog
• data processing

isoc servers
xroot

     The following servers are needed to allow processing of new data (older data on fermi-xrd or HPSS will not be available):

• fermi-gpfs01/02, fermi-gpfs05/06 and fermi-gpfs07/08
  These are the servers that make up the fermi xrootd gpfs space.  Each pair servers a part of the total gpfs space.
  • fermi-gpfs05 runs the xrootd server for the gpfs space
• fermilnx-v02 (redirector)
• fermilnx-v12 (redirector)

nfs

• Pretty much everything that's currently on staas-gpfs50/51
• Parts of the non-HA Fermi NFS file system

LSF

• 50 hosts should let us keep up (inlcuding ASP)

Here's what ISOC tasks need:

FASTCopy chain
--------------
staas-gpfs50/51
fermilnx01
fermilnx02
trscron
fermilnx-v03 (Archiver)
Whatever the pipeline server runs on.
xroot servers
astore-new system (HPSS)

Web servers
-----------
tomcat01 Commons
tomcat06 rm2
tomcat09 Pipeline-II
tomcat10 FCWebView, ISOCLogging, MPWebView
TelemetryMonitor, TelemetryTableWebUI
tomcat11 DataProcessing
tomcat12 TelemetryTrending

Science Pipelines

Gravitational Wave analysis (Nicola)

• Runs once per GW event reported from Global GW detectors
• Large variability in CPU requirement due to varying size of GW localization in sky
• Estimate 300 core-hours per day per GW event (e.g., 10 hours on 30 cores)
  →This would be 4 hequ hosts for about 10 hours per GW event

Flare Advocate analysis

• Batch jobs submitted to follow up on flare alert, typically once/day at most
• Batch job runs ~30 minutes on ~100 cores?
  →About 12 hequ-class nodes are needed for this analysis

Burst Analysis

• Batch job(s) submitted to follow up on gamma-ray burst detection
• Six jobs/burst, medium queue, rhel6
• Recent 7-day week had 11 triggers, so >1/day
  →Six hequ batch nodes should cover this need

FAVA (Fermi All-Sky Variability Analysis)

• Runs weekly
• Can postpone routine FAVA analysis until after the outage

High availability racks

For general information about the High-availability racks, Shirley provided this pointer to the latest list:

"Service Now, Knowledge Base,  search for "High Availability" , following link for current servers"

And here is the current statement about high-availability functionality:

Current Services in HA Racks
•CATER application
•Confluence application
•Data center management tool
•Drupal web
•Email lists
•Email transport infrastructure
•ERP application
•Exchange email
•EXO application
•Facilities monitoring
•Fermi application
•IT Ticketing system
•Network infrastructure
•Site Security infrastructure
•Unix authentication infrastructure
•Unix AFS infrastructure
•Unix mailboxes
•Unix monitoring
•VPN 
•Windows authentication infrastructure
•Windows file servers and SAN
•Windows monitoring
•Windows web

Supporting documentation

Email from Steve Tether with some storage-related information:

Change "fermilnx01 or fermilnx02" to "fermilnx01 and fermilnx02". While services can all be shifted to one of those machines, frankly it's a pain.

The partition staas-cnfs50lb:/gpfs/slac/ha/fs1/g/fermi/u23 currently has 554 GB free. This is where we store:
    - Incoming FASTCopy packages (L0 data, HSK data).
    - Outgoing FASTCopy packages (L1 data, mission planning).
    - Unpacked LAT raw data (L0, HSK, etc.)

FASTCopy packages for both L0 and L1 data are archived daily to "astore-new" and are then deleted within 24 hours. "astore-new" is a POSIX-compliant filesystem interface to HPSS that replaced the old "astore" interface. This is driven by the old GLAST Disk Archiver service.  The packages are also archived to xrootd daily.  Unpacked raw data is also archived to xrootd but is retained for 60 days on u23. The unpacked raw data on xrootd is a "live" backup in the sense that it can be accessed by ISOC tools and L1 reconstruction if needed, though that option is not normally enabled.

We get something like 16 GB of L0 data daily. If archiving to astore-new is turned off then we would have to retain the original incoming L0 FC packages, the unpacked L0 data and the L1 FC packages. Naively assuming that all of these to be about the same size that means retaining 48GB or more per day so u23 would fill up in 11.5 days or less. And we'd probably start experiencing problems as it approached being 100% full.

If the astore-new archiving were kept going but the xrootd archiving were suspended, then we would retain only the 16 GB of unpacked L0 data per day  which would fill up u23 in 30 days or so.

So I would recommend changing the classification of  "astore (non-Fermi server)" from NC to XC for this long of an outage.  And rename "astore" to "astore-new (HPSS)". I see that the Archiver server fermilnx-v03 is already classified as XC, so that's good.

The partition staas-cnfs50lb:/gpfs/slac/ha/fs1/g/fermi/u41 is used by the halfpipe to store events extracted from LAT raw data. The events would take up 16 GB daily times some modest expansion factor. That partition needs to be kept going for normal processing. I don't know how long the event data is retained but the partition currently has 4.4 TB free so it shouldn't be a problem in any event.

All the rest of the page seems OK.

Wilko's statement regarding space currently available in xrootd:

Nicola's estimate of batch power needed for GW follow up pipeline:

Stefano's comment on Flare Advocates:

Sara's response to Flare advocate script question:

They are usually submitted to the “medium” queue, in general relatively fast to finish (~30min) and not that much demanding resources wise, roughly a hundred cores I’d say. If you need more detailed info I do not know them on top of my head, there were also some changes lately to the code (I believe).  I or some of the FA coord. managing this would need to take a look into the code. E.g. you were asking about the batch node, what other info may help?

Teddy adds: "The FA scripts tend to be run each day at most."

Dan's statement on various Pipelines (including FAVA):

We have a few analysis pipelines that currently use the batch system. These include the burst advocate analysis, the gravitational wave followup, and FAVA.  The gravitational wave analysis typically requires thousands of jobs to be launched to analyze a large portion of the sky, so I think it’s probably hopeless to keep that up during the outage. FAVA runs on weekly timescales, so we can probably safely catch up that analysis once the batch farm comes back to full strength. The burst advocate analysis gets launched a little more than once a day.  Counting up the past week, we had 11 triggers in 7 days.  Each trigger launches 6 jobs and each job goes to the medium queue using rhel6. 

I can take the appropriate steps to deactivate the gravitational wave followup analysis and FAVA leading up to the outage. Let me know if you think we’d be able to keep the burst advocate analysis running and I’ll take the appropriate actions.  

Brian's proposal to move all VMs to H.A.:

Juliyana's report on VMs in H.A.:

Currently Fermi has these hypervisors and VMS in HA and Non-HA

HA:
Fermilnx-v06/v07/08
Fermilnx01/02/04

NON-HA:
Fermi-ci-test01
Fermilnx-v02/v03/v04/v05/v10/v11/v12/v13/v14/v15/v16/v17/v18/v19

---

These are fermi hypervisors. There are four of them. Two in HA and two in Non-HA.

HA: fermi-vmclust03 and fermivmclust04
Non-HA: fermi-vmclust01 and fermi-vmclust02

Excerpt from Christian Pama's spreadsheet on H.A. rack contents:

      fermi-vmclust03
      fermi-vmclust04
      fermilnx03-vmm
      fermi-oracle03
      wain031
      staas-gpfs50
      staas-gpfs51

      sca-oracle01
      scalnx02-vmm
      scalnx10-vmm01
      scalnx11-vmm02
      scalnx12-vmm01
      scalnx13-vmm02

Current tomcat server VMs:

Other VM's:

Gotchas from the Dec 2017 outage

1. We did not specify the "medium" LSF queue in our requirements, but ended up needing it
2. There was a delay in getting all 50 hequ's operating (some were/are on H.A., but others are not)
3. The xrootd redirector had a problem and needed a restart