Data Partitioning Techniques for Partially Protected Caches to Reduce Soft Error Induced FailuresOutlineMotivationSoft Errors on an IncreaseMost Vulnerable CachesUnequal Data ProtectionPPC – Partially Protected CachesData Partitioning in a PPCSlide 9Our SolutionVulnerable TimeVulnerability and Failure RateData Partitions using VulnerabilityGoal of Data PartitioningDPExplore – Data Partitioning HeuristicsSlide 16Experimental SetupEvaluationExperimental ResultsSignificant Reduction of VulnerabilityMin Overheads of Energy and RuntimeSlide 22DPExplore vs. MC and GASlide 24Slide 25ConclusionThanks!Backup SlidesSoft Errors on IncreaseSlide 30Slide 31Experimental Setup for Page FailuresImpact of Page Partitions to a PPCVulnerability under No Runtime PenaltyEnergy and Runtime under No PenaltyCopyright © 2008 UCI ACES Laboratory http://www.cecs.uci.edu/~acesKyoungwoo Lee1, Aviral Shrivastava2, Nikil Dutt1, and Nalini Venkatasubramanian1Data Partitioning Techniques for Partially Protected Caches to Reduce Soft Error Induced Failures1Department of Computer ScienceUniversity of California at Irvine2Department of Computer Scienceand EngineeringArizona State UniversityCopyright © 2008 UCI ACES Laboratory http://www.cecs.uci.edu/~acesOutlineMotivation and Problem StatementOur SolutionExperimentsConclusionDIPES 08 #2Copyright © 2008 UCI ACES Laboratory http://www.cecs.uci.edu/~acesMotivationSoft errors threaten the reliability of the systemSoft errors are expected to increase by several orders of magnitude beyond sub-micron technologyExponential increase of soft error rate as technology scales [Hazucha, 00]Redundancy techniques incur high overheads of power and performanceTMR (Triple Modular Redundancy) exceeds 200% overheads without optimization [Nieuwland, 06]ECC (Error Correction Codes) incurs overheads of performance by 95% [Li, 05] and power by 22% in caches [ARM, 03]PPC (Partially Protected Caches) [Lee, 06] is promising for multimedia applicationsNo obvious solutions to partition data into a PPC for general applicationsDIPES 08 #3Copyright © 2008 UCI ACES Laboratory http://www.cecs.uci.edu/~acesTransistorSoft Errors on an IncreaseSER increases exponentially as technology scalesIntegration, voltage scaling, altitude, latitude015 hours MTTF1 month MTTFBit Flip[Baumann, 05]•MTTF: Mean time To FailureDIPES 08 #4Copyright © 2008 UCI ACES Laboratory http://www.cecs.uci.edu/~acesMost Vulnerable CachesCaches are most hit due to:Larger portion in processors (more than 50%) No masking effect (e.g., no logical masking)DIPES 08 #5Intel Itanium II ProcessorCopyright © 2008 UCI ACES Laboratory http://www.cecs.uci.edu/~acesUnequal Data ProtectionAll pages are not equally failure critical(e.g.) Multimedia data is failure non-critical(e.g.) Program variables are failure criticalFailures: system crash, infinite loop, segmentation faults, etcDIPES 08 #6Only 9 pages out of 83 are failure criticalCopyright © 2008 UCI ACES Laboratory http://www.cecs.uci.edu/~acesPPC – Partially Protected CachesPPC architectures provide an unequal protection for mobile multimedia systems [Lee, 06]Unprotected cache and Protected cache at the same level of memory hierarchyProtected cache is typically smaller to keep power and delay the same as or less than those of Unprotected cacheVery efficient in terms of power and performanceDIPES 08 #7UnprotectedCacheProtectedCacheProtectedCacheMemoryPPCProcessorPipelineCopyright © 2008 UCI ACES Laboratory http://www.cecs.uci.edu/~acesData Partitioning in a PPC Multimedia ApplicationsMultimedia data is failure non-critical  Map multimedia data into the unprotected cache in a PPCAll other data is failure critical  Map all other data into the protected cache in a PPCGeneral ApplicationsNo obvious partitioning existsThis limits the applicability of the PPCProblem StatementFind data partitions for a PPC to minimize the overheads of power and performance with maximal reliabilityDIPES 08 #8UnprotectedCacheProtectedCacheProtectedCacheMemoryPPCCopyright © 2008 UCI ACES Laboratory http://www.cecs.uci.edu/~acesOutlineMotivation and Problem StatementOur SolutionExploitation of Vulnerability to Partition DataData Partitioning HeuristicsExperimentsConclusionDIPES 08 #9Copyright © 2008 UCI ACES Laboratory http://www.cecs.uci.edu/~acesOur SolutionData Partitioning Techniques – DPExploreDesign space exploration using Vulnerability metric rather than failure ratesJust one evaluation (vulnerability) vs. hundreds simulations (failure rate)Efficient explorations compared to Exhaustive Search or Genetic AlgorithmData partitioning for general applicationsNow PPC is effective not only for multimedia applications but also for general applicationsDIPES 08 #10Copyright © 2008 UCI ACES Laboratory http://www.cecs.uci.edu/~acesVulnerable TimeVulnerable timeIt is vulnerable for the time when eventually data is read by CPU or written back to MemoryVulnerability of a PageSum of vulnerable times of data in a pagePage is of 1 KB data in our studyDIPES 08 #11ReadWriteEvictionIncomingdatat0t1t2t3VulnerableVulnerableInvulnerableo Soft errors between t0 and t1 (t2 and t3) can cause failures of applications – data is vulnerable between t0 and t1 (t2 and t3)o Soft errors between t1 and t2 do not cause failures of applications since data will be updated by CPU – data is invulnerable between t1 and t2Copyright © 2008 UCI ACES Laboratory http://www.cecs.uci.edu/~acesVulnerability and Failure RateVulnerable time closely estimates failure rateDIPES 08 #12Copyright © 2008 UCI ACES Laboratory http://www.cecs.uci.edu/~acesData Partitions using VulnerabilityPages causing high vulnerable time are failure critical (FC)They are mapped into the Protected Cache in a PPCOthers are failure non-critical (FNC) mapped into the Unprotected CacheDIPES 08 #13Processor PipelineProcessorUnprotectedCacheProtectedCacheProtectedCacheMemoryPPCFCPagesFNCPagesFNCFCCopyright © 2008 UCI ACES Laboratory http://www.cecs.uci.edu/~acesGoal of Data PartitioningMust be careful when partitioning pagesToo many pages onto the (smaller) protected cache incurs many misses causing high overheads Goal of data partitionsdiscovers interesting pages to be mapped into a PPCfinds the best