Supporting Nested Transactional Memory in LogTM Authors Michelle J Moravan Mark Hill Jayaram Bobba Ben Liblit Kevin Moore Michael Swift Luke Yen David Wood Presented By Shweta Padubidri

OVERVIEW Common Terms LogTM LogTM with nested Transactions Experiment Conclusion

Introduction Transactional Memory – Concurrency control mechanism to control access to shared memory in concurrent computing – Allows a group of load and store instructions in an atomic way Nested Transactional Memory – a new transaction started by an instruction already inside an existing transaction – Transaction Isolation: Changes made by nested transactions are not seen by the 'host' until it is committed.

Transactional Memory System Transactional Memory Systems Differences – (Data) Version Management Eager: record old values “elsewhere”; update “in place” Lazy: update “elsewhere”; keep old values “in place” – (Data) Conflict Detection Eager: detect conflict on every read/write Lazy: detect conflict at end (commit/abort)

Log Based TM Version Management – New values stored in place (in main memory) – Old values stored in a thread-private transaction log C Data BlockVA 00 R W c LogPtr TMcount LogBase 1000

Log Based TM

Version Management – When a commit occurs the LogTM processor clears its cache’s W bits and resets the thread’s log pointer – Drawback : aborts are slow – Abort: LogTM must undo the transaction by writing old values back – Reset the R/W bits – Undo proceeds from end of the log to the beginning (LIFO)

Example

Log Based TM Conflict Detection – Requesting processor sends a coherence request to the directory – Directory responds and forwards the request to one or more processors – Each responding processor examines local state to detect conflict – Responding processors either ack(no conflict) or nack(conflict) the request – Requesting processor resolves any conflict

Log Based TM

Nested LogTM Motivation – Facilitating Software Composition Closed Nesting – transactions remain isolated until parent commits – Partial Aborts improve performance – Enhancing Concurrency Open Nesting – transactions commit independent of parent – Manages isolation at a higher level of abstraction

Nested LogTM Motivation (Contd) – Escaping to Non-Transactional Systems Supports calls to lower level non-transactional system ( eg: OS), from within a transaction with escape actions.

Version Management in Nested LogTM Closed Nesting – Flat Nested transactions “flattened” into a single transaction Only outermost begins/commits are meaningful Any conflict aborts to outermost transaction – Partial Rollback Child transaction can be aborted independently Can avoid costly re-execution of parent transaction But child merges transaction state with parent on commit Most conflicts with child end up affecting the parent

Version Management in Nested LogTM Flat LogTM’s log is a single frame (Fixed Size header + undo records) Nested LogTM’s log is a stack of frames A frame contains: – Header (including saved registers and pointer to parent’s frame) – Undo records (block address, old value pairs) – Garbage headers (headers of committed closed transactions) – Commit action records – Compensating action records

Version Management in Nested LogTM The Log Structure LogPtr TM countHeader Undo record LogBase LogFrame Header 21

Closed Nested Commit Merge child’s log frame with parent’s – Mark child’s header as a “garbage header” – Copy pointer from child’s header to LogFrame LogFrame LogPtr TM count Undo record Header Undo record 1 2 Header

Nested LogTM Flat LogTM detects conflicts with directory coherence and Read (R ) and Write (W ) bits in caches Nested LogTM replicates R/W bits for each level Flash-Or circuit merges child and parent R/W bits

Example

Concurrency Higher-level atomicity – Child’s memory updates not undone if parent aborts – Use compensating action to undo the child’s forward action at a higher-level of abstraction Higher-level isolation – Release memory-level isolation – Programmer enforce isolation at higher level (e.g., locks) – Use commit action to release isolation at parent commit

Open Nesting Commit Actions – Execute in FIFO order when innermost open ancestor commits Compensating Actions – Discard when innermost open ancestor commits – Execute in LIFO order when ancestor aborts

Open Nesting Condition Condition O1: An open nested child transaction never modifies a memory location that has been modified by any ancestor. // initialize to 0 counter = 0; transaction_begin(); // top-level 1 counter++; // counter gets 1 open_begin(); // level 2 counter++; // counter gets 2 open_commit(abort_action(counter--));... // Abort and run compensating action // Expect counter to be restored to 0... transaction_commit(); // not executed

Open Nesting in LogTM Version Management – Open commit pops the most recent frame off the log – (Optionally) add commit and compensating action records – Compensating actions are run by the software abort handler Conflict Detection – R/W bits cleared on open commit – No ‘OR’ of the R/W bits occur

Example

Escape Actions Escape actions are used to implement low-level functions (eg: exception handlers, debuggers and other run-time support) but are hidden from high level programmers Low level escapes to software that is non- transactional (no atomicity or isolation) May not invoke a transaction May register commit and compensating actions Have no effect on the enclosing transactions It increases the variety of code that execute within transactions

Escape Action Rules Escape Actions need to Obey – Condition X1: No writes to Data Written by Ancestors – Conditions X2: No Writes to Data Accessed by Others – Condition X3: No Reads to Data Written by Others

Escape Action in LogTM Version Management – No changes are made to the log. Read and Write access coherent memory to return the value from the latest uncommitted write Conflict Detection – No changes are made to the Read and Write bits Implements escape actions with a per thread flag, which disables logging and conflict detection when set. If an ancestor transaction aborts while the escape action is running, abort is delayed until the flag is cleared.

Experiment BTree Micro benchmark – Each Thread makes a repeated access to a shared tree – Lookup Probability(85%) and Insert(15%) – Each insert or lookup is executed as a top level transaction – New nodes are allocated from a shared free list using pen and closed nested transaction

B-Tree: Closed Nesting

B-Tree: Open Nesting

Conclusion Closed Nesting (partial rollback) – Easy to implement--segment the transaction log (stack of log frames)/Replicate R & W bits – Small performance gains for LogTM Open Nesting – Easy to implement--software abort handling allows easy execution of commit actions and compensating actions – Big performance gains Escape Actions – Provide non-transactional operations inside transactions

Questions ???