1. Efficient software checkpointing framework for speculative techniques
ECE Connections 2006
Co-Supervisors: Prof. Greg. Steffan
Prof. Cristiana Amza
Chuck (Chengyan) Zhao
Department of Computer Science
University of Toronto
Jun. 09, 2006
Might need to introduce my supervisors to the audience

2. Chip Multi-Processor (CMP) is now everywhere
IBM:
Power 4
Power 5
Intel:
Montecito
Smithfield
AMD:
dual-core Opteron, Athlon X2
Four-core Opteron
Sun:
UltraSparc T1: 32 cores
UltraSparc T2: 64 cores
Sony, Toshiba, IBM:
Cell:9 cores
… …
Power 4
Dual-core Intel chip
We are interested in improving the performance of a single application, using the abundant CMP resources (which most of would stay idle most of the time)
Dual-core Opteron
Cell
use CMP for single-threaded applications through parallelization

3. Parallelization Techniques
Automatic Parallelization
conservative + precise: prove of non dependence
limited domain
Speculative Parallelization
non-conservative
has to recover from failures
focus: speculative parallelization, use TLS

4. Thread-Level Speculation (TLS) Parallelism
Code example
for ( …){ …
*p = …;
… = … *q; }
difficult to parallelize automatically
uncertain dependence between *p and *q
might be runtime or user-input dependent
Points at slide while talking. 2. turn each loop iteration into a thread 3.
checkpointing scheme + dependence testing

5. How Thread-Level Speculation works
TLS
…*q
*p…  
violation 
Recover 
…*q
Exec.
Time
We take a sequential program and carve it into threads. Watch for violations
We then execute the threads speculatively in parallel.
The speculative part is that we don’t know whether these threads are actually independent.
Instead, we depend on runtime support to tell us whether the threads actually were independent
whenever we have violated a data dependence we simply re-execute that thread so that it is redone with the proper value, otherwise we can commit the speculative work.
But even when speculation has failed, we can still reduce overall execution time
by exploiting the available parallelism.
If you are usually right, then it is faster to apologize when you are wrong than to always ask for permission
exploit available thread-level parallelism

6. Memory Checkpointing Compiler Transformations
mark region of interest
backup each memory write (store)
generate buffer refresh calls
generate recovery code
remove region marking delimiters
start_instrument();
setjmp(buf1);
for(…){
refresh_ckpt();
backup_mem(a);
a = …;
backup_mem(b);
b = …; … }
if(error_spec()){ ckp_restore();
longjmp(buf1); }
stop_instrument();
mention those function calls are currently organized into a runtime library

7. Preliminary Results: MCF in SPEC2KINT
index
fname 1
refresh_potential() 2
bea_compute_red_cost() 3
primal_bea_mpp() 4
1 + 2 5
1 + 3 6
2 + 3 7
Picked SPEC2000 CPU INT Benchmark suite (make 10 / 12 applications available)
Remember to show the key point: performance degradation is can be up to 50%, but have large room of improvements

8. Challenges and Future Work
Challenges: software overhead
Proposed Solutions: optimizations
inlining
optimal buffer sizing and refreshing placement
memory optimizations
Applications
value prediction
debugging support
reliability enhancement
TLS (long term)
...
Mention that the challenge of software-only checkpointing is to significantly reducing the software overhead by aggressively optimizations

9. Questions and Answers