An HPC abstraction over MPI that uses pipes and pydash primitives. Blazer will handle all the MPI orchestration behind the scenes for you. You just work strictly with data and functions. Easy!
From pypi
$ pip install blazer
From clone
$ git clone https://github.com/radiantone/blazer
$ cd blazer
$ make init installNOTE: For some tests ensure you have slurm configured properly (single or muli-machine). However, using slurm is not required to use blazer.
$ make lint
(venv) $ mpirun -n 2 python setup.py test blazer/tests/test_parallel.py::test_parallel PASSED [ 50%] blazer/tests/test_pipeline.py::test_pipeline PASSED [100%] ============================== 2 passed in 0.48s =============================== ctrl-c
or
$ ./bin/tests.sh
import blazer from blazer.hpc.mpi import parallel, pipeline, partial as p, scatter, where, select, filter, rank, size def calc_some(value, *args): """ Do some calculations """ result = { 'some': value } return result def calc_stuff(value, *args): """ Do some calculations """ result = { 'this': value } return result def add_date(result): from datetime import datetime if type(result) is dict: result['date'] = str(datetime.now()) return result def calc_more_stuff(result): """ Do some more calculations """ if type(result) is list: result += [{'more':'stuff'}] elif type(result) is dict: result['more'] = 'stuff' return result INPUT_DATA = 'that' with blazer.begin(): result1=parallel([ p(calc_stuff, 1), p(calc_stuff, 2), p(calc_stuff, 3), p(calc_stuff, 4), p(calc_stuff, 5) ]) blazer.print("PARALLEL1:",result1) if blazer.ROOT: r = list( result1 | where(lambda g: where(lambda g: g['this'] > 1)) | select(lambda g: p(calc_stuff, g['this']*2)) ) # Run the composed computation in parallel, wait for result result = parallel(r) blazer.print("PARALLEL2:",result) r=pipeline([ p(calc_stuff, 'DATA'), p(pipeline, [ calc_some, add_date ]), calc_stuff ]) blazer.print("PIPELINE:",r) scatter_data = scatter(list(range(0,(size*2)+2)), calc_some) blazer.print("SCATTER_DATA:",scatter_data) result = pipeline([ p(calc_stuff, INPUT_DATA), add_date, scatter_data, p(parallel,[ calc_some, p(pipeline,[ calc_stuff, p(parallel, [ calc_some, calc_some ]), calc_stuff ]), calc_some ]), calc_more_stuff ]) blazer.print("PIPELINE RESULT:",result) def get_data(): """ Data generator """ for i in range(0,(size*2)): yield i result = scatter(get_data(), calc_some) blazer.print("SCATTER:",result)
To run:
(venv) $ export PYTHONPATH=.
(venv) $ mpirun -n 4 python blazer/examples/example1.py
PARALLEL1: [{'this': 1}, {'this': 2}, {'this': 3}, {'this': 4}, {'this': 5}]
PARALLEL2: [{'this': 4}, {'this': 6}, {'this': 2}, {'this': 8},; {'this': 10}]
PIPELINE: {'this': {'some': ({'this': 'DATA'},), 'date': '2022-02-11 02:47:23.356461'}}
SCATTER_DATA: [{'some': 0}, {'some': 1}, {'some': 2}, {'some': 3}, {'some': 4}, {'some': 5}, {'some': 6}, {'some': 7}, {'some': 8}, {'some': 9}, {'some': None}, {'some': None}]
PIPELINE RESULT: [{'this': [{'this': ([{'some': 0}, {'some': 1}, {'some': 2}, {'some': 3}, {'some': 4}, {'some': 5}, {'some': 6}, {'some': 7}, {'some': 8}, {'some': 9}, {'some': None}, {'some': None}],)}, {'some': {'some': [{'some': 0}, {'some': 1}, {'some': 2}, {'some': 3}, {'some': 4}, {'some': 5}, {'some': 6}, {'some': 7}, {'some': 8}, {'some': 9}, {'some': None}, {'some': None}]}}]}, {'some': 'some'}, {'more': 'stuff'}]
[0, 1, 2, 3, 4, 5, 6, 7]
SCATTER: [{'some': 0}, {'some': 1}, {'some': 2}, {'some': 3}, {'some': 4}, {'some': 5}, {'some': 6}, {'some': 7}]
A map/reduce example
import blazer from blazer.hpc.mpi import map, reduce def sqr(x): return x * x def add(x, y=0): return x+y with blazer.begin(): result = map(sqr, [1, 2, 3, 4]) blazer.print(result) result = reduce(add, result) blazer.print(result)
To run:
(venv) $ export PYTHONPATH=.
(venv) $ mpirun -n 4 python blazer/examples/example3.py
[1, 4, 9, 16]
30
Blazer supports the notion of streaming compute to handle jobs where the data can't fit into memory on a single machine.
In this example we implement a map/reduce computation where everything is streaming from the source data through the results without loading all the data into memory.
""" Streaming map/reduce example """ from itertools import groupby from random import randrange from typing import Generator import blazer from blazer.hpc.mpi import stream def datagen() -> Generator: for i in range(0, 1000): r = randrange(2) v = randrange(100) if r: yield {"one": 1, "value": v} else: yield {"zero": 0, "value": v} def key_func(k): return k["key"] def map(datum): datum["key"] = list(datum.keys())[0] return datum def reduce(datalist): from blazer.hpc.mpi import rank _list = sorted(datalist, key=key_func) grouped = groupby(_list, key_func) return [{"rank": rank, key: list(group)} for key, group in grouped] with blazer.begin(): import json mapper = stream(datagen(), map, results=True) reducer = stream(mapper, reduce, results=True) if blazer.ROOT: for result in reducer: blazer.print("RESULT", json.dumps(result, indent=4))
NOTE: blazer has (currently) only been tested on
mpirun (Open MPI) 4.1.0
Blazer is a high-performance computing (HPC) library that hides the complexities of a super computer's message-passing interface or (MPI). Users want to focus on their code and their data and not fuss with low-level API's for orchestrating results, building pipelines and running fast, parallel code. This is why blazer exists!
With blazer, a user only needs to work with simple, straightforward python. No cumbersome API's, idioms, or decorators are needed. This means they can get started quicker, run faster code, and get their jobs done faster!
Blazer is designed around the concept of computing primitives or operations. Some of the primitives include:
- parallel - For computing a list of tasks in parallel
- pipeline - For computing a list of tasks in sequence, passing the results along
- map - For mapping a task to a dataset
- reduce - For mapping a task to a data list and computing a single result
In addition there are other primitives to help manipulate lists of tasks or data, such as:
- where - Filter a list of tasks or data elements based on a function or lambda
- select - Apply a function to each list element and return the result
Blazer uses convenient context handlers to control blocks of code that need to be scheduled to MPI processes behind the scenes. There are two types of context handlers currently.
blazer.begin() is a mandatory context that enables the MPI scheduler behind the various primitives to operate correctly.
import blazer blazer.begin(): def get_data(): """ Data generator """ for i in range(0, (size * 2)): yield i result = scatter(get_data(), calc_some) blazer.print("SCATTER:", result)
blazer.gpu() is a context that requests (from the invisible MPI scheduler) dedicated access to a specific GPU on your MPI node fabric.
import logging import blazer import numpy as np from blazer.hpc.mpi.primitives import host, rank from numba import vectorize from timeit import default_timer as timer def dovectors(): @vectorize(['float32(float32, float32)'], target='cuda') def dopow(a, b): return a ** b vec_size = 100 a = b = np.array(np.random.sample(vec_size), dtype=np.float32) c = np.zeros(vec_size, dtype=np.float32) start = timer() dopow(a, b) duration = timer() - start return duration with blazer.begin(gpu=True): # on-fabric MPI scheduler with blazer.gpu() as gpu: # on-metal GPU scheduler # gpu object contains metadata about the GPU assigned print(dovectors())
Blazer supports synchronizing shared data across ranks seamlessly. Here is an example of sharing a dictionary where each rank adds its own data to the dictionary and it is available to all other ranks magically!
from random import randrange import blazer from blazer.hpc.mpi.primitives import rank with blazer.environment() as vars: rv = randrange(10) vars["rank" + str(rank)] = [ {"key": randrange(10)}, randrange(10), randrange(10), randrange(10), ] print("RANK:", rank, "DATA", vars.vars) blazer.print(vars["rank1"])
Blazer comes with a built-in design pattern for performing cross-cluster HPC. This is useful if you want to allocate compute resources on different super-computers and then build a pipeline of jobs across them. Here is a simple example using ALCF's Cooley and Theta systems (which are built into blazer).
from blazer.hpc.alcf import cooley, thetagpu from blazer.hpc.local import parallel, pipeline, partial as p # Log into each cluster using MFA password from MobilePASS cooleyjob = cooley.job(user='dgovoni', n=1, q="debug", A="datascience", password=True, script="/home/dgovoni/git/blazer/testcooley.sh").login() thetajob = thetagpu.job(user='dgovoni', n=1, q="single-gpu", A="datascience", password=True, script="/home/dgovoni/git/blazer/testthetagpu.sh").login() def hello(data, *args): return "Hello "+str(data) # Mix and match cluster compute jobs with local code tasks # in serial chaining cooleyjob("some data").then(hello).then(thetajob).then(hello) # Run a cross cluster compute job result = pipeline([ p(thetajob,"some data2"), p(cooleyjob,"some data1") ]) print("Done")
When each job .login() method is run, it will gather the MFA login credentials for that system and then use that to schedule jobs on that system via ssh.
Notice the use of the pipeline primitive above. It's the same primitive you would use to build your compute workflows! Composable tasks and composable super-computers.