- 1 Personal Information
- 2 Project Introduction
- 3 Improve Thread Safety
- 4 Multi-Threading Speedups
- 5 Performance Tweaks in Single Threaded Code
- 6 Time-Line
- 7 About Me
- 8 Miscellaneous
Name: Pulkit Mittal
IRC Nick: hoiji
STEP is the current standard for exchange of CAD data between different software packages. A STEP-File (also called Part21 file) is the most widely used data exchange form of STEP. Part21 files may be very huge (hundreds of Mb). These files are known to be heavily resource consuming (in terms of computing time or memory consumption) when being imported. The computing time taken by the processing of such files can be significantly reduced by using multi-thread programming techniques.
However before incorporating multi-threading into the STEP-Libraries we must ensure that the libraries are thread safe or else we might run into data races which are hard to debug leading a detrimental impact on development.
The Project can roughly divided into 3 major tasks. A detailed description of each is given in the following sections.
Improve Thread Safety: This is the first priority for this project. Starting with cllazyfile STEP library, I would introduce thread safety in the cleditor, clstepcore & cldai respectively in that order.
Introduce Multi-threading into the STEP Libraries: This task is highly dependent upon the first task. Once the concerned library has been deemed to be thread safe, we can start taking advantage of it to boost performance. As a proof of Concept, I intend to at-least one multi-threaded task.
Performance Tweaks in Single Threaded Code: This is a kind of a filler task. Since the 1st task would involve going through and understanding the code written in the aforementioned libraries. I would use that opportunity to point out, discuss and implement small tweaks which would increase the performance of the single threaded code.
Improve Thread Safety
The STEP Class Libraries consist of around 20K lines of code. I will first scan through the Documentation and ReadMe files and then review the code. This will help me figure out what resources could be possibly be shared by multiple threads. Once I have found the critical data / global variables, it would be easy to find critical sections and apply appropriate locking or other means of mutual exclusion. All major changes to the code will be documented.
Shared Resources: Apart from the std containers, the STEP Library also take dependency on resources (i.e data structures) defined in src/base/judy/* & src/clutils. In order to improve thread safety without hitting performance, some knowledge of the thread safety properties offered by the by the above data structures will be needed. Some examples of the utilization of such information are as follows.
- Judy Arrays: A Judy Array is a data structure that has high performance, low memory usage. They are extensively used in lazy loading. The lazy loading code itself is designed to help avoid using huge amounts of memory unless there is an actual need to map various properties in lazyInstMgr.h (Lazy Instance Manager). They are neither concurrent read safe nor concurrent write safe. In order to make a judy object concurrent write safe a mutex need to be allocated. The same can be done to make the judy object thread safe for concurrent read accesses. But this is bound to hit performance in cases where the Judy-Array is used as read-only object once it has been initialized. An alternative way to support concurrent read access would be by cloning the Judy object with judy_clone for use by each thread.
- STL-Containers: STL containers like vectors are used in various places inside the code. These are concurrent write unsafe but concurrent read safe. Thus, they will need mutexes similar to read-write locks.
Creating Unit Tests: Since all the existing units tests are for a single threaded code, I will create new tests which will spawn multiple threads to check the thread safety of libraries. This can be done by asking the threads would to read/write files in different threads simultaneously. These test will help me identify any deadlocks or data-races that might have crept in the code. Further, these test be of use to any future developers too.
Initial Progress and Findings
- Introduced myself and discussed about the project in the scl-dev mailing List and the #stepcode IRC channel.
- Forked the stepcode project on github to my handle and cloned it on my local machine.
- Gone through the source code of src/cllazyfile. I have read the appropriate documentation of the src/cleditor to get a fair idea of its functionality.
- Read about thread safe properties of judy arrays and std containers.
- Modify STEP Class Libraries for a more readable, maintainable and thread safe code.
- Modify the Documentation to reflect any major design changes
- Wherever required, Unit Tests will be made for checking thread safety.
Once I have finished part one of the project, I can start making STEP Libraries multi-threaded. Their could be possibly many ways of improving performance by taking advantage of multi-threading.
However as ensuring the thread safety can potentially take the majority of the GSOC duration, I plan to implement only one feature which would demonstrate the advantage of using multi-threading in lazy-loading. If time permits I can try adding other multi-threading features as well. For every feature I add there will be appropriate documentation and unit tests.
- A Module for Parsing and Processing data in different threads: Multi-threading can be used in Lazy-Loading by reading a file in one thread and storing the related data into in-memory structures using another thread. This approach might even improve performance on a single processor.
- (If Time Permits) A Module for Creating Multiple objects in Parallel: This module would take a list of objects as input and create multiple in-memory objects for each in parallel. This would especially increase performance on a multiprocessor leading to quicker loading.
- (If Time Permits) A Module for Loading Multiple Files in Parallel: This could benefit some upcoming STEP capabilities such as the use of External Element References to some entities (e.g. shape_aspect) in separate STEP files. So loading these files in parallel and supporting the references between them would increase performance of those applications that will take advantage of this capability.
Performance Tweaks in Single Threaded Code
This involves identifying inefficiencies in the code and and discussing them with my mentor. If I am not behind schedule I would also implement small patches which would remove the inefficiencies and increase performance.
This part of the project I plan to do concurrently with part one of my project. As the latter involves lot of reviewing of the code, this will increase my familiarity with the code. Also I intend to run and profile some already made tests which will help me identify the bottlenecks.
Initial Progress and Findings
- I have made a small modification the lazy-loading module. The code is currently under review. Once merged with the main code it is expected to give a 1.5x speedup to lazy-loading. The main idea for the modification was to buffer an already read data in one portion of the code.
Availability & Commitments
- Pre-GSOC: My college semester ends on 12th May. As this period would also require a good amount of commitment towards my college activities (tonnes of assignments, projects and exams). I will be able spend only around 10-20 hours a week on an average to prepare for my GSOC project.
- Early/Mid-GSOC: My summer vacations would last till 20th July. As I have no prior-commitments during this period, I will able to spend around 50-60 hrs per week on my GSOC project. In case I find that I am lagging behind, I am willing to spend up to 80 hrs per week for the project.
- Late-GSOC: After my college reopens on 21st July, up-till the GSOC pens down date i.e 18th August, I will be able to spend 40-50 hrs per week on the GSOC Project.
- Current - April, 15: Review the source code of src/cllazyfile, test the single threaded version it and start making it thread safe.
- April, 15 - April, 30: Thread safety for src/cllazyfile. (As mentioned before making Unit Test is implied)
- May, 1 - May, 12: College Exams!!
- May, 13 - May, 26: Thread safety for src/cleditor.
- May, 27 - June, 30: Thread safety for src/clstepcore. I may have to give a large time to this library as it is huge!
- July, 1 - July, 10: Thread safety for src/cldai.
- July, 11 - July, 31: Implement multi-threading in lazy-loading.
- August, 1 - August, 18: Buffer.
I am a 4th year Dual Degree CSE student studying in Indian Institute of Technology (IIT) Delhi, India (UTC+05:30). (Dual Degree is a 5 year course, in which students obtain their Bachelor’s as well as their Master’s degree in Technology in the area of Computer Science and Engineering.)
Programming: I have coded a lot in C, Java, Perl, Bash and PHP. I am also familiar with C++, C#, Python, SML, VHDL. I have used open-source softwares such as svn, gdb, llvm, gprof, qemu etc.
Interest Spectrum: My interest spectrum is quite large and varies from software systems (multi-threading, concurrency bug analysis, compiler optimizations, security) to hardware-software interface (multiprocessor programming, high performance architecture). I never miss a chance of making my program more efficient (in terms of computing time) and later brag about it!
Major Coding Experiences
Internship: In summer 2013, I had done an internship at Microsoft, India Development Center, Hyderabad. I was associated with the Windows Azure Integration Team. My team was impressed with my work and even extended me an offer of joining them as a Full-Time employee once I have completed my degree (i.e. in 2015).
Frankly, I was searching for a open source project in which I could put my experience with multi-threaded programming to practical use. Thats how I spotted the STEPCode project of ensuring thread safety and improve performance. Owing to the friendliness of my potential mentors, their quick and encouraging responses to my queries, I got started on my project in no time.
This STEPCode project nicely fits into the my interest domain. It gives me an opportunity to polish my multi-threading concepts and learn about the caveats associated with them. Also I believe the prospect of increasing performance of the code (via single thread tweaks, multiple threads) would increase my motivation further!
//Rest TO BE FILLED LATER