Evaluating the Cost of Atomic Operations on Modern Architectures M - PowerPoint PPT Presentation

spcl.inf.ethz.ch @spcl_eth A TOMICS : P ERFORMANCE D IMENSIONS Modified : in one cache and dirty Cache coherence state? Exclusive : in one cache and clean Shared : in >1 cache and clean

spcl.inf.ethz.ch @spcl_eth A TOMICS : P ERFORMANCE D IMENSIONS Modified : in one cache and dirty Cache coherence state? Exclusive : in one cache and clean Shared : in >1 cache and clean Invalid : garbage data

spcl.inf.ethz.ch @spcl_eth A TOMICS : P ERFORMANCE D IMENSIONS Architecture

spcl.inf.ethz.ch @spcl_eth A TOMICS : P ERFORMANCE D IMENSIONS Architecture

spcl.inf.ethz.ch @spcl_eth A TOMICS : P ERFORMANCE D IMENSIONS Architecture

spcl.inf.ethz.ch @spcl_eth A TOMICS : P ERFORMANCE D IMENSIONS Architecture

spcl.inf.ethz.ch @spcl_eth A TOMICS : P ERFORMANCE D IMENSIONS Architecture

spcl.inf.ethz.ch @spcl_eth R ESEARCH Q UESTIONS

spcl.inf.ethz.ch @spcl_eth R ESEARCH Q UESTIONS How do we model the performance of atomics?

spcl.inf.ethz.ch @spcl_eth R ESEARCH Q UESTIONS What is the How do we model the performance performance of difference between atomics? various atomics?

spcl.inf.ethz.ch @spcl_eth R ESEARCH Q UESTIONS What is the How do we model the performance performance of difference between atomics? various atomics? What is the influence of various parameters and mechanisms?

spcl.inf.ethz.ch @spcl_eth L ATENCY M ODEL Core … Cache Cache Cache Cache line

spcl.inf.ethz.ch @spcl_eth L ATENCY M ODEL Core … Cache Cache Cache Cache line

spcl.inf.ethz.ch @spcl_eth L ATENCY M ODEL Core Read for ownership … Cache Cache Cache Cache line

spcl.inf.ethz.ch @spcl_eth L ATENCY M ODEL Cache coherence state Core Read for ownership … Cache Cache Cache Cache line

spcl.inf.ethz.ch @spcl_eth L ATENCY M ODEL Cache coherence state Core Read for ownership = max(read latency, invalidation latency) … Cache Cache Cache Cache line

spcl.inf.ethz.ch @spcl_eth L ATENCY M ODEL Cache coherence state Core Read for ownership Cache line = max(read latency, invalidation latency) … Cache Cache Cache Cache line

spcl.inf.ethz.ch @spcl_eth L ATENCY M ODEL Cache coherence state Core Read for ownership Cache line = max(read latency, invalidation latency) … Cache Cache Cache Cache line

spcl.inf.ethz.ch @spcl_eth L ATENCY M ODEL Cache coherence state Core Read for ownership Cache line = max(read latency, invalidation latency) Execute = constant … Cache Cache Cache Cache line

spcl.inf.ethz.ch @spcl_eth L ATENCY M ODEL Cache coherence state Core Read for ownership Cache line = max(read latency, invalidation latency) Execute = constant … Cache Cache Cache Cache line Atomic

spcl.inf.ethz.ch @spcl_eth L ATENCY M ODEL Cache coherence state Core Read for ownership Cache line = max(read latency, invalidation latency) Execute = constant … Cache Cache Cache Cache line Atomic

spcl.inf.ethz.ch @spcl_eth L ATENCY M ODEL Cache coherence state Core Read for ownership Cache line = max(read latency, invalidation latency) Execute = constant … Cache Cache Cache Cache line Atomic Atomic

spcl.inf.ethz.ch @spcl_eth L ATENCY M ODEL Cache coherence state Core Read for ownership Cache line = max(read latency, invalidation latency) Execute = constant … Cache Cache Cache Cache line Atomic Cache Atomic coherence state

spcl.inf.ethz.ch @spcl_eth L ATENCY M ODEL E XCLUSIVE OR M ODIFIED S TATE Core Read for ownership Cache line = max(read latency, invalidation latency) Execute … Cache Cache Cache = constant Cache line Cache Atomic coherence state

spcl.inf.ethz.ch @spcl_eth L ATENCY M ODEL E XCLUSIVE OR M ODIFIED S TATE Core Read for ownership Cache line = read latency Execute … Cache Cache Cache = constant Cache line Cache Atomic coherence state

spcl.inf.ethz.ch @spcl_eth L ATENCY M ODEL E XCLUSIVE OR M ODIFIED S TATE mean of observed observed data predictions data Core Read for ownership Cache line = read latency Execute … Cache Cache Cache = constant Cache line Cache Atomic coherence state

spcl.inf.ethz.ch @spcl_eth L ATENCY H ASWELL , E XCLUSIVE

spcl.inf.ethz.ch @spcl_eth L ATENCY B ULLDOZER , E XCLUSIVE FAA CAS

spcl.inf.ethz.ch @spcl_eth L ATENCY H ASWELL , E XCLUSIVE Alignment?

spcl.inf.ethz.ch @spcl_eth L ATENCY Operand B ULLDOZER , E XCLUSIVE size? 64 bit 128 bit

spcl.inf.ethz.ch @spcl_eth B ANDWIDTH H ASWELL , A TOMICS

spcl.inf.ethz.ch @spcl_eth C ONCLUSIONS P ERFORMANCE I NSIGHTS

spcl.inf.ethz.ch @spcl_eth C ONCLUSIONS P ERFORMANCE I NSIGHTS The same latency of different atomics in most scenarios

spcl.inf.ethz.ch @spcl_eth C ONCLUSIONS P ERFORMANCE I NSIGHTS The same latency of different atomics in most scenarios CAS is the fastest for some cases

spcl.inf.ethz.ch @spcl_eth C ONCLUSIONS P ERFORMANCE I NSIGHTS Unaligned atomics should be avoided at all costs The same latency of different atomics in most scenarios CAS is the fastest for some cases

spcl.inf.ethz.ch @spcl_eth C ONCLUSIONS P ERFORMANCE I NSIGHTS Unaligned atomics should be avoided at all costs The same latency of different atomics in most scenarios No parallel execution (low bandwidth) even if there are no data deps CAS is the fastest for some cases

spcl.inf.ethz.ch @spcl_eth C ONCLUSIONS P ERFORMANCE I NSIGHTS Unaligned atomics should be avoided at all costs The same latency of different atomics in most scenarios No parallel execution (low bandwidth) even if there are no data deps CAS is the fastest for some cases Small operand sizes give best performance