Optimization and parallel computing
I have been optimizing and adding parallel processing to the system. Now it is literately 100x faster than I initially posted. And another few more times faster with parallel computing (although the performance does not scale ideally/linearly with the number of cores)
This is the testing environment that I’m using.
| Hardware | Info |
|---|---|
| Processor | AMD Ryzen 1700X (8 cores, 16 theads) @ 3.4GHz (Tubro OFF, locked at 3.4GHz for testing) |
| RAM | DDR4 2400MHz x2 |
| Operation System | Arch Linux x64 (kernel 4.20) |
| Compiler | GCC 8.2.1 |
| Parallel API | OpenMP |
Not boring everyone with details of optimization. I’ll just show the results.
Spatial Pooler
To test the performance of my HTM implementation. I decided to measure how long does TM/SP need to perform a specific task. For the SP, I measure how long it takes to generate and learn a 256 bit representation of the input SDR of different lengths with a potential pool ratio of 0.75.
The Spatial Pooler seems don’t scale well with the number of cores are there to solve the problem. There are only generally a 2.4x speedup for 1 core vs 4 cores. While Interestingly Hyper Thread / SMT threads seems not be doing and help. Using both 8 or 16 threads yields basically the same performance on my system (As I have 8 physical cores). This might indicate that some resource shared by the two cores has been used-up totally. I also found this behavior is also present on Intel processors.
I have also tested the performance of tiny-htm vs NuPIC.cpp… I’m ~10x slower in single thread inference speed. I totally don’t know why. Seems that NuPIC.cpp is unreasonably fast. Maybe I have set something wrong for NuPIC somewhere.
Temporal Memory
I did the same measurement for Temporal Memory. I measure how long a TM needs to infer and learn a random sequence of SDRs of different size. Sadly, I’m a few orders of magnitude slower than NuPIC.cpp in this case. I truly don’t know why.However, this time the HT/SMT threads does seem to help though. Running with 16 threads are noticeably (tho not much) faster than with 8 threads.
A 4x acceleration using 8 threads is good enough for me.
tiny-htm is very under-performing compared to NuPIC.cpp. But I hope this can serve as an reference to how HTM algorithm may behave under many threads.