1. Hardware Support for On-Demand Software Analysis Joseph L. Greathouse Advanced Computer Architecture Laboratory University of Michigan December 8, 2011

2. NIST: Software errors cost U.S. ~$60 billion/year FBI: Security Issues cost U.S. $67 billion/year  >⅓ from viruses, network intrusion, etc. Software Errors Abound 2

3. X = 0 Y=10 X = 5 Parallel Programming is Here – And it’s Hard! Hardware Plays a Role 3 x + y = 10 y + z = 20 Y = Z = 510 Y=10 10

4. Example of a Modern Bug 4 Thread 2 mylen=large Thread 1 mylen=small ptr ∅ Nov. 2010 OpenSSL Security Flaw if(ptr == NULL) { len=thread_local->mylen; ptr=malloc(len); memcpy(ptr, data, len); }

5. Example of a Modern Bug 5 if(ptr==NULL) memcpy(ptr, data2, len2) ptr LEAKED TIME Thread 2 mylen=large Thread 1 mylen=small ∅ len2=thread_local->mylen; ptr=malloc(len2); len1=thread_local->mylen; ptr=malloc(len1); memcpy(ptr, data1, len1)

6. Thread 2 mylen=large Thread 1 mylen=small if(ptr==NULL) len1=thread_local->mylen; ptr=malloc(len1); memcpy(ptr, data1, len1) len2=thread_local->mylen; ptr=malloc(len2); memcpy(ptr, data2, len2) Data Race Detection 6 Shared? Synchronized? TIME

7. Data Race Detection is Slow 7 PhoenixPARSEC

8. Inter-thread Sharing is What’s Important 8 if(ptr==NULL) len1=thread_local->mylen; ptr=malloc(len1); memcpy(ptr, data1, len1) len2=thread_local->mylen; ptr=malloc(len2); memcpy(ptr, data2, len2) Thread-local data NO SHARING Shared data, but NO DYNAMIC SHARING TIME

9. Very Little Dynamic Sharing 9 PhoenixPARSEC

10. Little Sharing Means Wasted Work 10 Multi-threaded Application Multi-threaded Application Software Race Detector Local Access Inter-thread sharing

11. Do Analysis On-Demand! 11 Multi-threaded Application Multi-threaded Application Software Race Detector Local Access Inter-thread sharing Inter-thread Sharing Monitor

12. Hardware Sharing Detector HITM in Cache Memory: W→R Data Sharing Hardware Performance Counters 12 Read Y S S I I S S I I HITM Core 1Core 2 Pipeline Cache 0 0 0 0 0 0 0 0 Perf. Ctrs 1 1 2 2 1 1 FAULT Write Y=5 M M Y=5

13. On-Demand Analysis on Real HW 13 Execute Instruction SW Race Detection Enable Analysis Disable Analysis HITM Interrupt? HITM Interrupt? Sharing Recently? Analysis Enabled? NO YES > 97% < 3%

14. Performance Difference 14 PhoenixPARSEC

15. Performance Increases 15 PhoenixPARSEC 51x Accuracy vs. Continuous Analysis: 97%

16. In Summary 16 Hardware makes constructing software difficult. Tools make software better. Hardware can (and should!) help these tools.

17. BACKUP SLIDES 17

18. Width Test 18