The growth in using GPUs for nongraphics
computation illustrates that
programmers are willing to redesign algorithms
[3B2-9] mmi2012030007.3d 17/5/012 12:42 Page 14
0
1
2
3
4
HT-H HT-L ATM CL BH CC AP
Normalized execution time
Ideal TM (Infinite TransWarp) Kilo TM (2 TransWarp/core)
Kilo TM (Infinite TransWarp) FG lock
8.1 6.2 5.7
Figure 7. Execution time of our applications with Kilo TM and fine-grained
locking (FG lock), normalized to the execution time of ideal TM. Lower is
better. (Infinite TransWarp: unlimited transaction concurrency; 2 TransWarp/
core: limiting each core to two concurrent transaction warps.)
....................................................................
14 IEEE MICRO
...............................................................................................................................................................................................
TOP PICKS
to exploit GPUs’ high computational efficiency.
Enhancing GPU architectures to
support robust synchronization mechanisms
such as transactional memory will lower the
risk of employing GPUs for more complex
applications. Some of the insights and
innovations gained from this work will likely
also apply to multicore CPUs as they scale to
support more concurrent threads. Our
evaluation with an ideal TM system motivates
us to make further improvements to
Kilo TM. Future work is also needed to
explore performance-tuning techniques for
GPU TM applications. Insights gained from
these explorations could contribute to a
GPU-optimized, intuitive TM interface that
doesn’t undermine programmers’ ability to
enhance application performance