-- The previous model was to enqueue individual map jobs (with a resolution of 1 map job per map call), to track the number of map calls submitted via a counter and a semaphore, and to use this information in each map job and reduce to control the number of map jobs, when reduce was complete, etc. All hideously complex.
-- This new model is vastly simply. The reducer basically knows nothing about the control mechanisms in the NanoScheduler. It just supports multi-threaded reduce. The NanoScheduler enqueues exactly nThread jobs to be run, which continually loop reading, mapping, and reducing until they run out of material to read, when they shut down. The master thread of the NS just holds a CountDownLatch, initialized to nThreads, and when each thread exits it reduces the latch by 1. The master thread gets the final reduce result when its free by the latch reaching 0. It's all super super simple.
-- Because this model uses vastly fewer synchronization primitives within the NS itself, it's naturally much faster at getting things done, without any of the overhead obvious in profiles of BQSR -nct 2.
-- reduceAsMuchAsPossible no longer blocks threads via synchronization, but instead uses an explicit lock to manage access. If the lock is already held (because some thread is doing reduce) then the thread attempting to reduce immediately exits the call and continues doing productive work. They removes one major source of blocking contention in the NanoScheduler
-- Created a separate, limited interface MapResultsQueue object that previously was set to the PriorityBlockingQueue.
-- The MapResultsQueue is now backed by a synchronized ExpandingArrayList, since job ids are integers incrementing from 0 to N. This means we avoid the n log n sort in the priority queue which was generating a lot of cost in the reduce step
-- Had to update ReducerUnitTest because the test itself was brittle, and broken when I changed the underlying code.
-- A few bits of minor code cleanup through the system (removing unused constructors, local variables, etc)
-- ExpandingArrayList called ensureCapacity so that we increase the size of the arraylist once to accommodate the upcoming size needs
-- Now each map job reads a value, performs map, and does as much reducing as possible. This ensures that we scale performance with the nct value, so -nct 2 should result in 2x performance, -nct 3 3x, etc. All of this is accomplished using exactly NCT% of the CPU of the machine.
-- Has the additional value of actually simplifying the code
-- Resolves a long-standing annoyance with the nano scheduler.
-- The NanoSchedule timing code (in NSRuntimeProfile) was crazy expensive, but never showed up in the profilers. Removed all of the timing code from the NanoScheduler, the NSRuntimeProfile itself, and updated the unit tests.
-- For tools that largely pass through data quickly, this change reduces runtimes by as much as 10x. For the RealignerTargetCreator example, the runtime before this commit was 3 hours, and after is 30 minutes (6x improvement).
-- Took this opportunity to improve the GATK ProgressMeter. NotifyOfProgress now just keeps track of the maximum position seen, and a separate daemon thread ProgressMeterDaemon periodically wakes up and prints the current progress. This removes all inner loop calls to the GATK timers.
-- The history of the bug started here: http://gatkforums.broadinstitute.org/discussion/comment/2402#Comment_2402
-- The previous nanoscheduler would deadlock in the case where an Error, not an Exception, was thrown. Errors, like out of memory, would cause the whole system to die. This bugfix resolves that issue
Doesn't actually fix the problem, and adds an unnecessary delay in closing down NanoScheduler, so reverting.
This reverts commit 66b820bf94ae755a8a0c71ea16f4cae56fd3e852.
-- Renamed TraversalErrorManager to the more general MultiThreadedErrorTracker
-- ErrorTracker is now used throughout the NanoScheduler. In order to properly handle errors, the work previously done by main thread (submit jobs, block on reduce) is now handled in a separate thread. The main thread simply wakes up peroidically and checks whether the reduce result is available or if an error has occurred, and handles each appropriately.
-- EngineFeaturesIntegrationTest checks that -nt and -nct properly throw errors in Walkers
-- Added NanoSchedulerUnitTest for input errors
-- ThreadEfficiencyMonitoring is now disabled by default, and can be enabled with a GATK command line option. This is because the monitoring doesn't differentiate between threads that are supposed to do work, and those that are supposed to wait, and therefore gives misleading results.
-- Build.xml no longer copies the unittest results verbosely
-- V3 + V4 algorithm for NanoScheduler. The newer version uses 1 dedicated input thread and n - 1 map/reduce threads. These MapReduceJobs perform map and a greedy reduce. The main thread's only job is to shuttle inputs from the input producer thread, enqueueing MapReduce jobs for each one. We manage the number of map jobs now via a Semaphore instead of a BlockingQueue of fixed size.
-- This new algorithm should consume N00% CPU power for -nct N value.
-- Also a cleaner implementation in general
-- Vastly expanded unit tests
-- Deleted FutureValue and ReduceThread
-- Now prints out a single combined NanoScheduler runtime profile report across all nano schedulers in use. So now if you run with -nt 4 you'll get one combined NanoScheduler profiler across all 4 instances of the NanoScheduler within TraverseXNano.
-- I've rewritten the entire NS framework to use a producer / consumer model for input -> map and from map -> reduce. This is allowing us to scale reasonably efficiently up to 4 threads (see figure). Future work on the nano scheduler will be itemized in a separate JIRA entry.
-- Restructured the NS code for clarity. Docs everywhere.
-- This is considered version 1.0
-- The NanoScheduler is doing a good job at tracking important information like time spent in map/reduce/input etc.
-- Can be disabled with static boolean in MicroScheduler if we have problems
-- See GSA-515 Nanoscheduler GSA-549 Retire TraverseReads and TraverseLoci after testing confirms nano scheduler version in single threaded version is fine
-- Separate updating cumulative traversal metrics from printing progress. There's now an updateCumulativeMetrics function and a printProgress() that only takes a current position
-- printProgress now soles relies on the time since the last progress to decide if it will print or not. No longer uses the number of cycles, since this isn't reliable in the case of nano scheduling
-- GenomeAnalysisEngine now maintains a pointer to the master cumulative metrics. getCumulativeMetrics never returns null, which was handled in some parts of the code but not others.
-- Update all of the traversals to use the new updateCumulativeMetrics, printProgress model
-- Added progress callback to nano scheduler. Every bufferSize elements this callback is invoked, allowing us to smoothly update the progress meter in the NanoScheduler
-- Rename MapFunction to NanoSchedulerMap and the same for reduce.
-- Helpful for understanding where the time goes to each bit of the code.
-- Controlled by a local static boolean, to avoid the potential overhead in general
-- TraverseReadsNano modified to read in all input data before invoking maps, so the input to TraverseReadsNano is a MapData object holding the sam record, the ref context, and the refmetadatatracker.
-- Update ValidateRODForReads to be tree reducible, using synchronized map and explicitly sort the output map from locations -> counts in onTraversalDone
-- Expanded integration tests to test nt 1, 2, 4.
-- TraversalReadsNano only creates the NanoScheduler once, and shuts it down onTraversalDone
-- Nicer debugging output in NanoScheduler
-- ReadShard has a getBufferSize() method now
-- Groups inputs for each thread so that we don't have one thread execution per map() call
-- Added shutdown function
-- Documentation everywhere
-- Code cleanup
-- Extensive unittests
-- At this point I'm ready to integrate it into the engine for CPU parallel read walkers
– Write general NanoScheduler framework in utils.threading. Test with reading via iterator from list of integers, map is int * 2, reduce is sum. Should be efficiency using resources to do sum of 2 * (sum(1 - X)).
Done!
CPU parallelism is nano threads. Pfor across read / map / reduce. Use work queue to implement.
Create general read map reduce framework in utils. Test parallelism independently before hooking up to Locus iterator
Represent explicitly the dependency graph. Scheduler should choose the work units that are ready for computation, that are marked as "completing a computation", and then finally that maximize the number of sequent available work units. May be worth measuring expected cost for read read / map / reduce unit and use it to balance the compute
As input is single threaded just need one thread to populate inputs, which runs as fast as possible on parallel pushing data to fixed size queue. Each push creates map job and links to upcoming reduce job.
Note that there's at most one thread for IO tasks, and all of the threads can contribute to CPU tasks
Mauricio Carneiro
Mark DePristo
David Roazen
Menachem Fromer