Condor Project Computer Sciences Department University of Wisconsin-Madison Using New Features in Condor 7.2
Outline › Startd Hooks › Job Router › Job Router Hooks › Power Management › Dynamic Slot Partitioning › Concurrency Limits › Variable Substitution › Preemption Attributes 2
Startd Job Hooks › Users wanted to take advantage of Condor’s resource management daemon (condor_startd) to run jobs, but they had their own scheduling system. Specialized scheduling needs Jobs live in their own database or other storage rather than a Condor job queue 3
Our solution › Make a system of generic “hooks” that you can plug into: A hook is a point during the life-cycle of a job where the Condor daemons will invoke an external program Hook Condor to your existing job management system without modifying the Condor code 4
How does Condor communicate with hooks? › Passing around ASCII ClassAds via standard input and standard output › Some hooks get control data via a command-line argument (argv) › Hooks can be written in any language (scripts, binaries, whatever you want) so long as you can read/write Stdin/out 5
What hooks are available? › Hooks for fetching work (startd): FETCH_JOB REPLY_FETCH EVICT_CLAIM › Hooks for running jobs (starter): PREPARE_JOB UPDATE_JOB_INFO JOB_EXIT 6
HOOK_FETCH_JOB › Invoked by the startd whenever it wants to try to fetch new work FetchWorkDelay expression › Stdin: slot ClassAd › Stdout: job ClassAd › If Stdout is empty, there’s no work 7
HOOK_REPLY_FETCH › Invoked by the startd once it decides what to do with the job ClassAd returned by HOOK_FETCH_WORK › Gives your external system a chance to know what happened › argv[1]: “accept” or “reject” › Stdin: slot and job ClassAds › Stdout: ignored 8
HOOK_EVICT_CLAIM › Invoked if the startd has to evict a claim that’s running fetched work › Informational only: you can’t stop or delay this train once it’s left the station › Stdin: both slot and job ClassAds › Stdout: ignored 9
HOOK_PREPARE_JOB › Invoked by the condor_starter when it first starts up (only if defined) › Opportunity to prepare the job execution environment Transfer input files, executables, etc. › Stdin: both slot and job ClassAds › Stdout: ignored, but starter won’t continue until this hook exits › Not specific to fetched work 10
HOOK_UPDATE_JOB_INFO › Periodically invoked by the starter to let you know what’s happening with the job › Stdin: slot and job ClassAds Job ClassAd is updated with additional attributes computed by the starter: ImageSize, JobState, RemoteUserCpu, etc. › Stdout: ignored 11
HOOK_JOB_EXIT › Invoked by the starter whenever the job exits for any reason › Argv[1] indicates what happened: “exit”: Died a natural death “evict”: Booted off prematurely by the startd (PREEMPT == TRUE, condor_off, etc) “remove”: Removed by condor_rm “hold”: Held by condor_hold 12
HOOK_JOB_EXIT … › “HUH!?! condor_rm? What are you talking about?” The starter hooks can be defined even for regular Condor jobs, local universe, etc. › Stdin: copy of the job ClassAd with extra attributes about what happened: ExitCode, JobDuration, etc. › Stdout: ignored 13
Defining hooks › Each slot can have its own hook ”keyword” Prefix for config file parameters Can use different sets of hooks to talk to different external systems on each slot Global keyword used when the per-slot keyword is not defined › Keyword is inserted by the startd into its copy of the job ClassAd and given to the starter 14
Defining hooks: example # Most slots fetch work from the database system STARTD_JOB_HOOK_KEYWORD = DATABASE # Slot4 fetches and runs work from a web service SLOT4_JOB_HOOK_KEYWORD = WEB # The database system needs to both provide work and # know the reply for each attempted claim DB_DIR = /usr/local/condor/fetch/db DATABASE_HOOK_FETCH_WORK = $(DB_DIR)/fetch_work.php DATABASE_HOOK_REPLY_FETCH = $(DB_DIR)/reply_fetch.php # The web system only needs to fetch work WEB_DIR = /usr/local/condor/fetch/web WEB_HOOK_FETCH_WORK = $(WEB_DIR)/fetch_work.php 15
Semantics of fetched jobs › Condor_startd treats them just like any other kind of job: All the standard resource policy expressions apply (START, SUSPEND, PREEMPT, RANK, etc). Fetched jobs can coexist in the same pool with jobs pushed by Condor, COD, etc. Fetched work != Backfill 16
Semantics continued › If the startd is unclaimed and fetches a job, a claim is created › If that job completes, the claim is reused and the startd fetches again › Keep fetching until either: The claim is evicted by Condor The fetch hook returns no more work 17
Limitations of the hooks › If the starter can’t run your fetched job because your ClassAd is bogus, no hook is invoked to tell you about it We need a HOOK_STARTER_FAILURE › No hook when the starter is about to evict you (so you can checkpoint) Can implement this yourself with a wrapper script and the SoftKillSig attribute 18
Job Router › Automated way to let jobs run on a wider array of resources Transform jobs into different forms Reroute jobs to different destinations 19
What is “job routing”? 20 Universe = “vanilla” Executable = “sim” Arguments = “seed=345” Output = “stdout.345” Error = “stderr.345” ShouldTransferFiles = True WhenToTransferOutput = “ON_EXIT” Universe = “grid” GridType = “gt2” GridResource = \ “cmsgrid01.hep.wisc.edu/jobmanager-condor” Executable = “sim” Arguments = “seed=345” Output = “stdout” Error = “stderr” ShouldTransferFiles = True WhenToTransferOutput = “ON_EXIT” JobRouter Routing Table: Site 1 … Site 2 … final status routed (grid) joboriginal (vanilla) job
Routing is just site-level matchmaking › With feedback from job queue number of jobs currently routed to site X number of idle jobs routed to site X rate of recent success/failure at site X › And with power to modify job ad change attribute values (e.g. Universe) insert new attributes (e.g. GridResource) add a “portal” grid proxy if desired 21
Configuring the Routing Table › JOB_ROUTER_ENTRIES list site ClassAds in configuration file › JOB_ROUTER_ENTRIES_FILE read site ClassAds periodically from a file › JOB_ROUTER_ENTRIES_CMD read periodically from a script example: query a collector such as Open Science Grid Resource Selection Service 22
Syntax › List of sites in new ClassAd format [ Name = “Grid Site 1”; … ] [ Name = “Grid Site 2”; … ] [ Name = “Grid site 3”; … ] … 23
Syntax [ Name = “Site 1”; GridResource = “gt2 gk.foo.edu”; MaxIdleJobs = 10; MaxJobs = 200; FailureRateThreshold = 0.01; JobFailureTest = other.RemoteWallClockTime < 1800 Requirements = target.WantJobRouter is True; delete_WantJobRouter = true; set_PeriodicRemove = JobStatus == 5; ] 24
What Types of Input Jobs? › Vanilla Universe › Self Contained (everything needed is in file transfer list) › High Throughput (many more jobs than cpus) 25
Grid Gotchas › Globus gt2 no exit status from job (reported as 0) › Most grid universe types must explicitly list desired output files 26
JobRouter vs. Glidein › Glidein - Condor overlays the grid job never waits in remote queue job runs in its normal universe private networks doable, but add to complexity need something to submit glideins on demand › JobRouter some jobs wait in remote queue (MaxIdleJobs) job must be compatible with target grid semantics simple to set up, fully automatic to run 27
Job Router Hooks › Truly transform jobs, not just reroute them E.g. stuff a job into a virtual machine (either VM universe or Amazon EC2) › Hooks invoked like startd ones 28
HOOK_TRANSLATE › Invoked when a job is matched to a route › Stdin: route name and job ad › Stdout: transformed job ad › Transformed job is submitted to Condor 29
HOOK_UPDATE_JOB_INFO › Invoked periodically to obtain extra information about routed job › Stdin: routed job ad › Stdout: attributes to update in routed job ad 30
HOOK_JOB_FINALIZE › Invoked when routed job has completed › Stdin: ads of original and routed jobs › Stdout: modified original job ad or nothing (no updates) 31
HOOK_JOB_CLEANUP › Invoked when original job returned to schedd (both success and failure) › Stdin: Original job ad › Use for cleanup of external resources 32
Power Management › Hibernate execute machines when not needed › Condor doesn’t handle waking machines up yet › Information to wake machines available in machine ads 33
Configuring Power Management › HIBERNATE Expression evaluated periodically by all slots to decide when to hibernate All slots must agree to hibernate › HIBERNATE_CHECK_INTERVAL Number of seconds between hibernation checks 34
Setting HIBERNATE › HIBERNATE must evaluate to one of these strings: “NONE”, “0” “S1”, “1”, “STANDBY”, “SLEEP” “S2”, “2” “S3”, “3”, “RAM”, “MEM” “S4”, “4”, “DISK”, “HIBERNATE” “S5”, “5”, “SHUTDOWN” › These numbers are ACPI power states 35
Power Management on Linux › On linux, theses methods are tried in order for setting power level: pm-UTIL tools /sys/power /proc/ACPI › LINUX_HIBERNATION_METHOD can be set to pick a favored method 36
Sample Configuration ShouldHibernate = \ ((KeyboardIdle > $(StartIdleTime)) \ && $(CPUIdle) \ && ($(StateTimer) > (2 * $(HOUR))) HIBERNATE = ifThenElse( \ $(ShouldHibernate), “RAM”, “NONE” ) HIBERNATE_CHECK_INTERVAL = 300 LINUX_HIBERNATION_METHOD = “/proc” 37
Dynamic Slot Partitioning › Divide slots into chunks sized for matched jobs › Readvertise remaining resources › Partitionable resources are cpus, memory, and disk 38
How It Works › When match is made… New sub-slot is created for job and advertised Slot is readvertised with remaining resources › Slot can be partitioned multiple times › Original slot ad never enters Claimed state But may eventually have too few resources to be matched › When claim on sub-slot is released, resources are added back to original slot 39
Configuration › Resources still statically partitioned between slots › SLOT_TYPE_ _PARTITIONABLE Set to True to enable dynamic partition within indicated slot 40
New Machine Attributes › In original slot machine ad PartitionableSlot = True › In ad for dynamically-created slots DynamicSlot = True › Can reference these in startd policy expressions 41
Job Submit File › Jobs can request how much of partitionable resources they need request_cpus = 3 request_memory = 1024 request_disk =
Dynamic Partitioning Caveats › Cannot preempt original slot or group of sub-slots Potential starvation of jobs with large resource requirements › Partitioning happens once per slot each negotiation cycle Scheduling of large slots may be slow 43
Concurrency Limits › Limit job execution based on admin- defined consumable resources E.g. licenses › Can have many different limits › Jobs say what resources they need › Negotiator enforces limits pool-wide 44
Concurrency Example › Negotiator config file MATLAB_LIMIT = 5 NFS_LIMIT = 20 › Job submit file concurrency_limits = matlab,nfs:3 This requests 1 Matlab token and 3 NFS tokens 45
New Variable Substitution › $$(Foo) in submit file Existing feature Attribute Foo from machine ad substituted › $$([Memory * 0.9]) in submit file New feature Expression is evaluated and then substituted 46
More Info For Preemption › New attributes for these preemption expressions in the negotiator… PREEMPTION_REQUIREMENTS PREEMPTION_RANK › Used for controlling preemption due to user priorities 47
Preemption Attributes › Submitter/RemoteUserPrio User priority of candidate and running jobs › Submitter/RemoteUserResourcesInUse Number of slots in use by user of each job › Submitter/RemoteGroupResourcesInUse Number of slots in use by each user’s group › Submitter/RemoteGroupQuota Slot quota for each user’s group 48
Thank You! › Any questions? 49