Extended Memory Semantics - Persistent shared object memory and parallelism for Node.js and Python
OSX | Linux | Node 4.1-14.x, Python2/3:
EMS makes possible persistent shared memory parallelism between Node.js, Python, and C/C++.
Extended Memory Semantics (EMS) unifies synchronization and storage primitives to address several challenges of parallel programming:
A modern multi-core server has 16-32 cores and nearly 1TB of memory, equivalent to an entire rack of systems from a few years ago. As a consequence, jobs formerly requiring a Map-Reduce cluster can now be performed entirely in shared memory on a single server without using distributed programming.
Inter-language example in interlanguage.{js,py} The animated GIF demonstrates the following steps:
EMS extends application capabilities to include transactional memory and
other fine-grained synchronization capabilities.
EMS implements several different parallel execution models:
|
|
|
|
EMS operations may performed using any JSON data type, read-modify-write operations may use any combination of JSON data types. like operations on ordinary data.
Atomic read-modify-write operations are available in all concurrency modes, however collectives are not available in user defined modes.
Atomic Operations: Read, write, readers-writer lock, read when full and atomically mark empty, write when empty and atomically mark full
Primitives: Stacks, queues, transactions
Read-Modify-Write: Fetch-and-Add, Compare and Swap
Collective Operations: All basic OpenMP collective operations are implemented in EMS: dynamic, block, guided, as are the full complement of static loop scheduling, barriers, master and single execution regions
Map-Reduce is often demonstrated using word counting because each document can be processed in parallel, and the results of each document's dictionary reduced into a single dictionary. This EMS implementation also iterates over documents in parallel, but it maintains a single shared dictionary across processes, atomically incrementing the count of each word found. The final word counts are sorted and the most frequently appearing words are printed with their counts.
The performance of this program was measured using an Amazon EC2 instance:
c4.8xlarge (132 ECUs, 36 vCPUs, 2.9 GHz, Intel Xeon E5-2666v3, 60 GiB memory
The leveling of scaling around 16 cores despite the presence of ample work
may be related to the use of non-dedicated hardware:
Half of the 36 vCPUs are presumably HyperThreads or otherwise shared resource.
AWS instances are also bandwidth limited to EBS storage, where our Gutenberg
corpus is stored.
A benchmark similar to STREAMS
gives us the maximum speed EMS double precision
floating point operations can be performed on a
c4.8xlarge (132 ECUs, 36 vCPUs, 2.9 GHz, Intel Xeon E5-2666v3, 60 GiB memory
.
Transactions and Work Queues Example
Transactional performance is measured alone, and again with a separate
process appending new processes as work is removed from the queue.
The experiments were run using an Amazon EC2 instance:
c4.8xlarge (132 ECUs, 36 vCPUs, 2.9 GHz, Intel Xeon E5-2666v3, 60 GiB memory
Six EMS arrays are created, each holding 1,000,000 numbers. During the benchmark, 1,000,000 transactions are performed, each transaction involves 1-5 randomly selected elements of randomly selected EMS arrays. The transaction reads all the elements and performs a read-modify-write operation involving at least 80% of the elements. After all the transactions are complete, the array elements are checked to confirm all the operations have occurred.
The parallel process scheduling model used is block dynamic (the default),
where each process is responsible for successively smaller blocks
of iterations. The execution model is bulk synchronous parallel, each
processes enters the program at the same main entry point
and executes all the statements in the program.
forEach
loops have their normal semantics of performing all iterations,
parForEach
loops are distributed across threads, each process executing
only a portion of the total iteration space.
Immediate Transactions: Each process generates a transaction on integer data then immediately performs it. |
Transactions from a Queue: One of the processes generates the individual transactions and appends them to a work queue the other threads get work from. Note: As the number of processes increases, the process generating the transactions and appending them to the work queue is starved out by processes performing transactions, naturally maximizing the data access rate. |
Immediate Transactions on Strings: Each process generates a transaction appending to a string, and then immediately performs the transaction. |
Elem. Ref'd: Total number of elements read and/or written Table Updates: Number of different EMS arrays (tables) written to Trans. Performed: Number of transactions performed across all EMS arrays (tables) Trans. Enqueued: Rate transactions are added to the work queue (only 1 generator thread in these experiments) |
EMS internally stores tags that are used for synchronization of user data, allowing synchronization to happen independently of the number or kind of processes accessing the data. The tags can be thought of as being in one of three states, Empty, Full, or Read-Only, and the EMS intrinsic functions enforce atomic access through automatic state transitions.
The EMS array may be indexed directly using an integer, or using a key-index mapping from any primitive type. When a map is used, the key and data itself are updated atomically.
EMS memory is an array of JSON values (Number, Boolean, String, Undefined, or Object) accessed using atomic operators and/or transactional memory. Safe parallel access is managed by passing through multiple gates: First mapping a key to an index, then accessing user data protected by EMS tags, and completing the whole operation atomically. |
EMS Data Tag Transitions & Atomic operations: F=Full, E=Empty, X=Don't Care, RW=Readers-Writer lock (# of current readers) CAS=Compare-and-Swap, FAA=Fetch-and-Add |
For a more complete description of the principles of operation, contact the author at [email protected]
Because all systems are already multicore, parallel programs require no additional equipment, system permissions, or application services, making it easy to get started. The reduced complexity of lightweight threads communicating through shared memory is reflected in a rapid code-debug cycle for ad-hoc application development.
To build and test all C, Python 2 and 3, and Node.js targets, a makefile can automate most build and test tasks.
dunlin> make help
Extended Memory Semantics -- Build Targets
===========================================================
all Build all targets, run all tests
node Build only Node.js
py Build both Python 2 and 3
py[2|3] Build only Python2 or 3
test Run both Node.js and Py tests
test[_js|_py|_py2|_py3] Run only Node.js, or only Py tests, respectively
clean Remove all files that can be regenerated
clean[_js|_py|_py2|_py3] Remove Node.js or Py files that can be regenerated
EMS is available as a NPM Package. EMS depends on the Node addon API (N-API) package.
npm install ems
Download the source code, then compile the native code:
git clone https://github.com/mogill/ems.git
cd ems
npm install
Python users should download and install EMS git (see above). There is no PIP package, but not due lack of desire or effort. A pull request is most welcome!
Click here for Detailed Examples.
On a Mac and most Linux distributions EMS will "just work", but some Linux distributions restrict access to shared memory. The quick workaround is to run jobs as root, a long-term solution will vary with Linux distribution.
Run the work queue driven transaction processing example on 8 processes:
npm run <example>
Or manually via:
cd Examples
node concurrent_Q_and_TM.js 8
Running all the tests with 8 processes:
npm run test # Alternatively: npm test
cd Tests
rm -f EMSthreadStub.js # Do not run the machine generated script used by EMS
for test in `ls *js`; do node $test 8; done
As of 2016-05-01, Mac/Darwin and Linux are supported. A windows port pull request is welcomed!
EMS 1.0 uses Nan for long-term Node.js support, we continue to develop on OSX and Linux via Vagrant.
EMS 1.3 introduces a C API.
EMS 1.4 Python API
EMS 1.4.8 Improved examples and documentation
EMS 1.5 Refactored JS-EMS object conversion temporary storage
EMS 1.6 [This Release] Updated to replace deprecated NodeJS NAN API with the N-API.
EMS 1.7 [Planned] Key deletion that frees all resources. Replace open hashing with chaining.
EMS 1.8 [Planned] Memory allocator based on R. Marotta, M. Ianni, A. Scarselli, A. Pellegrini and F. Quaglia, "NBBS: A Non-Blocking Buddy System for Multi-core Machines," 2019 19th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID), Larnaca, Cyprus, 2019, pp. 11-20, doi: 10.1109/CCGRID.2019.00011.
EMS 1.9 [Planned] Vectorized JSON indexer.
EMS 1.10 [Planned] Support for persistent main system memory (PMEM) when multiple processes are supported.
EMS 2.0 [Planned] New API which more tightly integrates with ES6, Python, and other dynamically typed languages languages, making atomic operations on persistent memory more transparent.
BSD, other commercial and open source licenses are available.
Jace A Mogill specializes in hardware/software co-design of resource constrained computing at both the largest and smallest scales. He has over 20 years experience with distributed, multi-core, FPGA, CGRA, GPU, CPU, and custom computer architectures.