ATLAS Site Availability and Performance (ASAP)
Since December 2014, a new way to measure ATLAS computing availability has been introduced. It's called ASAP / ATLAS Site Availability and Performance. From what I have gleaned about it, ASAP is a metric to replace ADCD site status. The status of the "PandaResource" of an analysis queue is used by ASAP (production queues may be included in future). A site is considered to be unavailable when its analysis queue is in test mode. This document will briefly describe some of the implications of this.
Various tests are sent out to sites using a framework called "HammerCloud". The results of various tests are combined using an algorithm. The HammerCloudATLASOperations wiki page describes it all. The upshot is that once a set of tests have failed, the site is set to test, i.e. not available. Ultimately that is used by ASAP to determine the site's unavailability periods.
How to get the important alerts
Unfortunately, at the moment, when a queue is set to test mode, the notification email is sent to cloud support and doesn’t necessarily go to to our site admins. Some site admins may be able to subscribe themselves to the list (email@example.com) at e-groups.cern.ch. Admins without the necessary security credentials can request to be subscribed; ask Elena Korolkova, Alessandra Forti or another GridPP representative of ATLAS.
Once you are getting the alerts, it's usually easy to set up filters that can find the messages for your site by searching the subject field for the name of the site's queues. A list of all Panda queues can be found here: dashb-atlas-ssb.cern.ch.
Where to check ASAP site status
- A new monitor will be prepared in grafana
Once you have had an alert or otherwise found out that you have a problem, it's time to look into the cause. The HammerCloud Site Overview is a good place to begin. Select your site and some times to examine, and choose functional tests (FTs used for blacklisting).You'll get some tables representing the PandaResources (queues) at your site. I can't understand the naming convention, but the important queue at present follows this pattern: ANALY_<site_mnemonic>_SL6 , e.g. for Liverpool, the relevant one is ANALY_LIV_SL6. Each grid has tabs along the top that represent the test type (template). You need to check them all. Look for tests that are f (red) or m (orange). Those are the ones that have caused trouble. Clicking on the test takes you to the summary. You use the information in this page, and the ones linked from it, to debug the problem. The first thing to note is the Job info field from the second table (Job List). If the Job list table has several entries, choose jobs with status of "failed".
How to use HC detailed status records to debug common scenarios
A website called bigpanda is the primary source of information. Click on PanDA resource name from the list of the left (e.g. ANALY_LIV_SL6) and on the next page click on "View: ... jobs" to see the full list of jobs, or "View: ... job errors" to limit the output to jobs that went wrong. It's your own choice where to go next, but the links in the "Site error summary" table bring up a summary table for one job. Clicking on a job's PandaID gives you access to all the job's log files, stdout, stderr etc.
An online tutorial shows how to use big panda monitor. Please note that the BigPandaMonitor is presently incomplete and lacks some user interface controls. Nonetheless, it is useful and the tutorial gives clues about how to navigate it.
Sometimes you strike it lucky, the the job info field tells you exactly what's wrong. For example, PanDA ID 2346079525 has this entry:
lcg-cp: error while loading shared libraries: libglobus_ftp_client.so.2:
Obviously there's a bad library version on the worker node.
As often as not, a lot more digging is required to find the cause. The following sections outline various approaches to different issues.
More Obscure Errors