Object Duplication for Improving Reliability

Transcription

1 Object Duplication for Improving Reliability G. Chen, G. Chen, M. Kandemir, N. Vijaykrishnan, M. J. Irwin Department of Computer Science and Engineering The Pennsylvania State University

2 Soft Errors S Current G D A particle strike n+ n channel n p substrate B Soft errors or transient errors are circuit errors caused due to excess charge carriers induced primarily by external radiations

3 Soft Errors Why soft errors become problematic? As CMOS device sizes decrease, the charge stored at each node decreases Potentially leads to a much higher rate of soft errors Low-power operating modes accentuates the soft error problem

4 Challenges for Soft Error Protection Balance several factors Cost Extra hardware required Flexibility Different requirements of multiple applications Overhead Performance, power, memory space Reliability

5 Java Virtual Machine (JVM) Java is widely used in embedded devices JVM has details about application s status and semantics error detection and recovery may be easier Heap is the dominant space consumer in JVM We focus on heap objects A B Free Space C Free Space

6 Checksum-based Schemes Error? A B C Checksum A B C Read Free Space Free Space A checksum for each field or each object Updated at write operations Checking is required for each read operation Error detection, but no error correction

7 Motivation for Object Duplication Average Heap Occupancy 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% auction calc firstaid image jpeg manyballs mvideo pushpuzzle AVERAGE Remaining heap can be used for other purposes, or example, storing object duplicates

8 Object Duplication Primary Area Duplicate Area D1 D2 D3 Free Space P3 P2 P1 Duplicate is updated on each write Duplicate is read when there is an error in the primary

9 Full Duplication Performance overhead Extra accesses on writes Read accesses when error happens Cache conflicts Performance overhead is very low Memory space requirement doubled Two ways to reduce memory space overhead Compression Selective duplication

10 Compression Based Duplication Primary Area Duplicate Area D1 D2 D3 Free Space P3 P2 P1 D1 D2 D3 Free Space P3 P2 P1

11 Compression Based Duplication All objects have compressed duplicates Performance overhead Compression on writes Decompression when primary copy corrupted Little overhead if the object is rarely updated Zero-removal compression/decompression Fast Java objects contains a lot of zero bytes

12 Life Cycle of an Object Write Read Drag Dead Unused Time Allocation First access Lifetime of object o Last reference removed Garbage Last access collected

13 Full Duplication Create duplicate Time range that duplicate exists Remove duplicate Drag Dead Allocation First access Lifetime of object o Time Last reference removed Garbage Last access collected

14 Selective Duplication Early termination based duplication Lazy duplication Allocation site based

15 Early Termination Based Create duplicate Remove duplicate Time range that duplicate exists Drag Dead Allocation First access Lifetime of object o Time Last reference removed Garbage Last access collected

16 Lazy Duplication Create duplicate Remove duplicate Time range that duplicate exists Drag Dead Allocation First access Lifetime of object o Time Last reference removed Garbage Last access collected

17 Allocation Site Based Based on profiling Sort allocation sites according to the access frequency of the object accessed Create duplicate at allocation sites leading to higher access frequency No duplicate creation at allocation sites leading to lower access frequency

18 Discussion of Selective Schemes Reliability is lower than full duplication Object life time Early termination based scheme favors shortliving objects Lazy duplication and allocation site based schemes favor long-living objects Static vs dynamic Early termination based and lazy duplication schemes are dynamic Allocation site based scheme is static

19 Experiment Setup Java Application KVM Shade MP Simulator Execution Cycles Memory Allocations Three types of statistics Heap space Error resilience Performance

20 Benchmarks Benchmark auction calc firstaid jpeg image manyballs mvideo pushpuzzle Description Ticket auction Calculator Firstaid information Jpeg viewer Photo album Bounding balls Video player Puzzle game

21 Allocated Heap Space auction calc CHK DUPL COMPDUPL DUPL increases the size of of the allocated heap space by a factor of of COMPDUPL reduces it it to to firstaid image jpeg manyballs mvideo pushpuzzle

22 350% 300% Heap Occupancy for Selective Schemes LAST-USE-100 LAST-USE-80 DUPL Heap Occupancy 250% 200% 150% 100% 50% LAST-USE-OPT COMPDUPL PRFL DELAYED Selective schemes performs well in in terms of of 0% heap space consumption. The ideal LAST-USE- OPT performs very well. PRFL and DELAYED are better than COMPDUPL. Time (K cycles)

23 Non-correctable Error Rates Percentage of Errors not Recovered 30% 25% 20% 15% 10% 5% 0% auction calc DUPL firstaid image jpeg COMPDUPL Average non-correctable error rate of of DUPL and COMPDUPL are 8.2% and 4.3%, respectively. manyballs mvideo pushpuzzle

24 Non-correctable Error Rates Percentage of Errors not recovered 80% 60% 40% 20% 0% auction No Duplicate Faulty Duplicate 98.3% 88.4% 89.1% calc Average non-correctable error rate of of LAST- USE-OPT, LAST-USE-100, LAST-USE-80, PRFL, From left to right: LAST-USE-OPT, LAST-USE-100, LASTand DELAYED USE-80, are 8.2%, PRFL, 8.2%, DELAYED. 50.2%, 19.1% and 41.6%, respectively. firstaid image jpeg manyballs mvideo pushpuzzle

25 Execution Cycles CHK DUPL COMPDUPL auction calc firstaid image jpeg Average execution cycles increase of of DUPL and COMPDUPL are 11.3% and 34.9%, respectively. manyballs mvideo pushpuzzle

26 Conclusion Object duplication improves data integrity and has good error coverage Compression reduces heap occupancy but increases execution cycles significantly Selective schemes help tradeoff heap space consumption, error recover rate, and performance

27 Thank you!