- StreamLEC is a stream processing system that specifically designed for stream machine learning applications and provides lightweight proactive fault tolerance via erasure coding. This source code showcases the prototype of StreamLEC and its usage.
- Zhinan Cheng, Lu Tang, Qun Huang, and Patrick P. C. Lee.
Enabling Low-Redundancy Proactive Fault Tolerance for Stream Machine Learning via Erasure Coding
Proceedings of the 40th International Symposium on Reliable Distributed Systems (SRDS 2021), September 2021.
- gcc and camke (version 2.6.4 or above), we have tested StreamLEC in Ubuntu 16.04 with gcc 5.4.0 and cmake 3.5.1.
- CPU and compiler support SIMD instruction AVX and AVX2.
- check the CPU flag avx and avx2 via
cat /proc/cpuinfo.
- check the CPU flag avx and avx2 via
- StreamLEC also relies on ZMQ and iniparser (with source code built-in).
- Step 1: Build StreamLEC core library.
- Create a directory in project root directory (e.g., release), create Makefile, and build the code
mkdir release && cd release cmake .. make -j- Set up the library.
sudo cp libstreamlec.so /usr/lib cd .. - Step 2: Compile example application.
- Provided applications
- Logistic regression training, without enabling hybrid coded computation
- Logistic regression prediction, enabling hybrid coded computation
- Here we take the streaming logistic regression training as example (check
./apps/logistic_regression/).- We train a logistic regression model in real time using the provided sample input
- use default encoder in source.
- use default decoder and implement
RecomputeandAggregatein the sink. - implement the
ProcessDatainterfaces in each processors
- We train a logistic regression model in real time using the provided sample input
- For compile
- use the provided Makefile.
cd apps/logistic_regression make -j
- Provided applications
- Step 3: Run example application
- Configurations
- one Source, one Sink, and three processors
- check the configuration file in
./apps/logistic_regression/sample.ini.- k = 2, r = 1, micro_batch size = 300.
- Input trace
- check
./apps/trace/sample.txtfor the example trace.- 10000 items, each with 12 attributes.
- check
- Run pseudo distributed mode using a single machine
- open one terminal, run the Source.
./source sample.ini Encoder- run 3 processors, each on a new terminal.
./processor sample.ini Processor1 ./processor sample.ini Processor2 ./processor sample.ini Processor3- open another terminal, run the sink
./sink sample.ini Decoder- check the output in each terminal to see the progress
- Immediately failure recovery
- You can kill any r processors (e.g., 1 in this case) during their running to verify the immediate recovery
- run in distributed mode
- Set up the core library on each node
- copy the core library
libstreamlec.soto the/usr/libdirectory of each node
- copy the core library
- change the configuration file
- update the ip address of Encoder, Decoder, and each Processor
- check the script in
./scriptto run StreamLEC- fill in the ip address of the cluster
- run
./setup.shto setup the preparation - run
./run_all.shto run an application- check the arguments for the script
- run
./stop_all.shto stop an application
- Set up the core library on each node
- Configurations
- users can check the provided interfaces and templates for building application in directory
./programming. - Source
- StreamLEC provides a default decoder based on RS code in the source (check
/source_interfaces.hpp), users have no need to modify the source code if use RS codes. - StreamLEC also allows users to deploy other erasure codes by re-implementing the
Encode()interfaces insource_interface.hpp).
- StreamLEC provides a default decoder based on RS code in the source (check
- Sink
- StreamLEC also provides a default decoder based on RS code in the sink (check
sink_interface.hpp), users only need to implement theRecomputeandAggregateinterfaces (check.sink_interface.hpp) for an applications. - Users also needs to re-implementing
Decode()interfaces if they want to uses a new erasure codes (checksink_interface.hpp).
- StreamLEC also provides a default decoder based on RS code in the sink (check
- Processor
- For processor, users only need to fill in the required interfaces, check
processor_interface.hpp.- if enable coded computation, users need to decouple the computation into two parts and fill in the
ProcessLinearandProcessNonlinear - otherwise only need to fill in the
ProcessDatainterface - fill in the
ProcessFeedbackinterfaces if there are feedbacks in the applications
- if enable coded computation, users need to decouple the computation into two parts and fill in the
- For processor, users only need to fill in the required interfaces, check
- Please contact [email protected] if you have any problems.