TaskMemoryManager

TaskMemoryManager manages the memory allocated to execute a single task (using MemoryManager).

TaskMemoryManager is created when TaskRunner is requested to run.

TaskMemoryManager
Figure 1. Creating TaskMemoryManager for Task

TaskMemoryManager assumes that:

  • The number of bits to address pages (aka PAGE_NUMBER_BITS) is 13

  • The number of bits to encode offsets in data pages (aka OFFSET_BITS) is 51 (i.e. 64 bits - PAGE_NUMBER_BITS)

  • The number of entries in the page table and allocated pages (aka PAGE_TABLE_SIZE) is 8192 (i.e. 1 << PAGE_NUMBER_BITS)

  • The maximum page size (aka MAXIMUM_PAGE_SIZE_BYTES) is 15GB (i.e. ((1L << 31) - 1) * 8L)

Creating Instance

TaskMemoryManager takes the following to be created:

TaskMemoryManager initializes the internal properties.

Spillable Memory Consumers

TaskMemoryManager tracks spillable memory consumers.

TaskMemoryManager registers a new memory consumer when requested to acquire execution memory.

TaskMemoryManager removes (clears) all registered memory consumer when requested to clean up all allocated memory.

Memory consumers are used to report memory usage when TaskMemoryManager is requested to show memory usage.

MemoryManager

TaskMemoryManager is given a MemoryManager when created.

TaskMemoryManager uses the MemoryManager for the following:

Cleaning Up All Allocated Memory

long cleanUpAllAllocatedMemory()

cleanUpAllAllocatedMemory clears page table.

FIXME

All recorded consumers are queried for the size of used memory. If the memory used is greater than 0, the following WARN message is printed out to the logs:

WARN TaskMemoryManager: leak [bytes] memory from [consumer]

The consumers collection is then cleared.

MemoryManager.releaseExecutionMemory is executed to release the memory that is not used by any consumer.

Before cleanUpAllAllocatedMemory returns, it calls MemoryManager.releaseAllExecutionMemoryForTask that in turn becomes the return value.

FIXME Image with the interactions to MemoryManager.
cleanUpAllAllocatedMemory is used exclusively when TaskRunner is requested to run (and cleans up after itself).

Acquiring Execution Memory

long acquireExecutionMemory(
  long required,
  MemoryConsumer consumer)

acquireExecutionMemory allocates up to required size of memory for the MemoryConsumer.

When no memory could be allocated, it calls spill on every consumer, itself including. Finally, acquireExecutionMemory returns the allocated memory.

acquireExecutionMemory synchronizes on itself, and so no other calls on the object could be completed.
MemoryConsumer knows its mode — on- or off-heap.

acquireExecutionMemory first calls memoryManager.acquireExecutionMemory(required, taskAttemptId, mode).

TaskMemoryManager is a mere wrapper of MemoryManager to track consumers?
FIXME

When the memory obtained is less than requested (by required), acquireExecutionMemory requests all consumers to release memory (by spilling it to disk).

acquireExecutionMemory requests memory from consumers that work in the same mode except the requesting one.

You may see the following DEBUG message when spill released some memory:

DEBUG Task [taskAttemptId] released [bytes] from [consumer] for [consumer]

acquireExecutionMemory calls memoryManager.acquireExecutionMemory(required, taskAttemptId, mode) again (it called it at the beginning).

It does the memory acquisition until it gets enough memory or there are no more consumers to request spill from.

You may also see the following ERROR message in the logs when there is an error while requesting spill with OutOfMemoryError followed.

ERROR error while calling spill() on [consumer]

If the earlier spill on the consumers did not work out and there is still memory to be acquired, acquireExecutionMemory requests the input consumer to spill memory to disk (that in fact requested more memory!)

If the consumer releases some memory, you should see the following DEBUG message in the logs:

DEBUG Task [taskAttemptId] released [bytes] from itself ([consumer])

acquireExecutionMemory calls memoryManager.acquireExecutionMemory(required, taskAttemptId, mode) once more.

memoryManager.acquireExecutionMemory(required, taskAttemptId, mode) could have been called "three" times, i.e. at the very beginning, for each consumer, and on itself.

It records the consumer in consumers registry.

You should see the following DEBUG message in the logs:

DEBUG Task [taskAttemptId] acquired [bytes] for [consumer]

acquireExecutionMemory is used when:

Allocating Memory Block for Tungsten Consumers

MemoryBlock allocatePage(
  long size,
  MemoryConsumer consumer)
It only handles Tungsten Consumers, i.e. MemoryConsumers in tungstenMemoryMode mode.

allocatePage allocates a block of memory (aka page) smaller than MAXIMUM_PAGE_SIZE_BYTES maximum size.

It checks size against the internal MAXIMUM_PAGE_SIZE_BYTES maximum size. If it is greater than the maximum size, the following IllegalArgumentException is thrown:

Cannot allocate a page with more than [MAXIMUM_PAGE_SIZE_BYTES] bytes

It then acquires execution memory (for the input size and consumer).

It finishes by returning null when no execution memory could be acquired.

With the execution memory acquired, it finds the smallest unallocated page index and records the page number (using allocatedPages registry).

If the index is PAGE_TABLE_SIZE or higher, releaseExecutionMemory(acquired, consumer) is called and then the following IllegalStateException is thrown:

Have already allocated a maximum of [PAGE_TABLE_SIZE] pages

It then attempts to allocate a MemoryBlock from Tungsten MemoryAllocator (calling memoryManager.tungstenMemoryAllocator().allocate(acquired)).

FIXME What is MemoryAllocator?

When successful, MemoryBlock gets assigned pageNumber and it gets added to the internal pageTable registry.

You should see the following TRACE message in the logs:

TRACE Allocate page number [pageNumber] ([acquired] bytes)

The page is returned.

If a OutOfMemoryError is thrown when allocating a MemoryBlock page, the following WARN message is printed out to the logs:

WARN Failed to allocate a page ([acquired] bytes), try again.

And acquiredButNotUsed gets acquired memory space with the pageNumber cleared in allocatedPages (i.e. the index for pageNumber gets false).

FIXME Why is the code tracking acquiredButNotUsed?

Another allocatePage attempt is recursively tried.

FIXME Why is there a hope for being able to allocate a page?

releaseExecutionMemory Method

void releaseExecutionMemory(long size, MemoryConsumer consumer)

releaseExecutionMemory…​FIXME

releaseExecutionMemory is used when:

getMemoryConsumptionForThisTask Method

long getMemoryConsumptionForThisTask()

getMemoryConsumptionForThisTask…​FIXME

getMemoryConsumptionForThisTask is used exclusively in Spark tests.

Displaying Memory Usage

void showMemoryUsage()

showMemoryUsage prints out the following INFO message to the logs (with the taskAttemptId):

Memory used in task [taskAttemptId]

showMemoryUsage requests every MemoryConsumer to report memory used. showMemoryUsage prints out the following INFO message to the logs for a MemoryConsumer with some memory usage (and excludes zero-memory consumers):

Acquired by [consumer]: [memUsage]

showMemoryUsage prints out the following INFO messages to the logs:

[amount] bytes of memory were used by task [taskAttemptId] but are not associated with specific consumers
[executionMemoryUsed] bytes of memory are used for execution and [storageMemoryUsed] bytes of memory are used for storage

showMemoryUsage is used when MemoryConsumer is requested to throw an OutOfMemoryError.

pageSizeBytes Method

long pageSizeBytes()

pageSizeBytes simply requests the MemoryManager for pageSizeBytes.

pageSizeBytes is used when…​FIXME

Freeing Memory Page — freePage Method

void freePage(MemoryBlock page, MemoryConsumer consumer)

pageSizeBytes simply requests the MemoryManager for pageSizeBytes.

pageSizeBytes is used when MemoryConsumer is requested to freePage and throwOom.

Getting Page — getPage Method

Object getPage(long pagePlusOffsetAddress)

getPage…​FIXME

getPage is used when…​FIXME

Getting Page Offset — getOffsetInPage Method

long getOffsetInPage(long pagePlusOffsetAddress)

getPage…​FIXME

getPage is used when…​FIXME

Logging

Enable ALL logging level for org.apache.spark.memory.TaskMemoryManager logger to see what happens inside.

Add the following line to conf/log4j.properties:

log4j.logger.org.apache.spark.memory.TaskMemoryManager=ALL

Refer to Logging.

Internal Properties

Name Description

acquiredButNotUsed

The size of memory allocated but not used.

allocatedPages

Collection of flags (true or false values) of size PAGE_TABLE_SIZE with all bits initially disabled (i.e. false).

TIP: allocatedPages is java.util.BitSet.

When allocatePage is called, it will record the page in the registry by setting the bit at the specified index (that corresponds to the allocated page) to true.

pageTable

The array of size PAGE_TABLE_SIZE with indices being MemoryBlock objects.

When allocating a MemoryBlock page for Tungsten consumers, the index corresponds to pageNumber that points to the MemoryBlock page allocated.

tungstenMemoryMode

MemoryMode (i.e. OFF_HEAP or ON_HEAP)

Set to the tungstenMemoryMode of the MemoryManager while TaskMemoryManager is created