Citi/scaler

Efficient, lightweight and reliable distributed computation engine.

Scaler provides a simple, efficient and reliable way to perform distributed computing using a centralized scheduler, with a stable and language-agnostic protocol for client and worker communications.

import math
from scaler import Client

with Client(address="tcp://127.0.0.1:2345") as client:
    # Submits 100 tasks
    futures = [
        client.submit(math.sqrt, i)
        for i in range(0, 100)
    ]

    # Collects the results and sums them
    result = sum(future.result() for future in futures)

    print(result)  # 661.46

Scaler is a suitable Dask replacement, offering significantly better scheduling performance for jobs with a large number of lightweight tasks while improving on load balancing, messaging and deadlocks.

Features

• Distributed computing across multiple cores and multiple servers
• Python reference implementation, with language-agnostic messaging protocol built on top of Cap'n Proto and ZeroMQ
• Graph scheduling, which supports Dask-like graph computing, with optional GraphBLAS support for very large graph tasks
• Automated load balancing, which automatically balances load from busy workers to idle workers, ensuring uniform utilization across workers
• Automated task recovery from worker-related hardware, OS or network failures
• Support for nested tasks, allowing tasks to submit new tasks
• top-like monitoring tools
• GUI monitoring tool

Installation

$ pip install scaler

# or with graphblas and uvloop support
$ pip install scaler[graphblas,uvloop]

Quick Start

Scaler operates around 3 components:

• A scheduler, responsible for routing tasks to available computing resources
• A set of workers, or cluster. Workers are independent computing units, each capable of executing a single task
• Clients running inside applications, responsible for submitting tasks to the scheduler.

Start local scheduler and cluster programmatically in code

A local scheduler and a local set of workers can be conveniently started using SchedulerClusterCombo:

from scaler import SchedulerClusterCombo

cluster = SchedulerClusterCombo(address="tcp://127.0.0.1:2345", n_workers=4)

...

cluster.shutdown()

This will start a scheduler with 4 task executing workers on port 2345.

Setting up a computing cluster from the CLI

The scheduler and workers can also be started from the command line with scaler_scheduler and scaler_cluster.

First, start the Scaler scheduler:

$ scaler_scheduler tcp://127.0.0.1:2345
[INFO]2023-03-19 12:16:10-0400: logging to ('/dev/stdout',)
[INFO]2023-03-19 12:16:10-0400: use event loop: 2
[INFO]2023-03-19 12:16:10-0400: Scheduler: monitor address is ipc:///tmp/127.0.0.1_2345_monitor
...

Then, start a set of workers (a.k.a. a Scaler cluster) that connect to the previously started scheduler:

$ scaler_cluster -n 4 tcp://127.0.0.1:2345
[INFO]2023-03-19 12:19:19-0400: logging to ('/dev/stdout',)
[INFO]2023-03-19 12:19:19-0400: ClusterProcess: starting 4 workers, heartbeat_interval_seconds=2, object_retention_seconds=3600
[INFO]2023-03-19 12:19:19-0400: Worker[0] started
[INFO]2023-03-19 12:19:19-0400: Worker[1] started
[INFO]2023-03-19 12:19:19-0400: Worker[2] started
[INFO]2023-03-19 12:19:19-0400: Worker[3] started
...

Multiple Scaler clusters can be connected to the same scheduler, providing distributed computation over multiple servers.

-h lists the available options for the scheduler and the cluster executables:

$ scaler_scheduler -h
$ scaler_cluster -h

Submitting Python tasks using the Scaler client

Knowing the scheduler address, you can connect and submit tasks from a client in your Python code:

from scaler import Client


def square(value: int):
    return value * value


with Client(address="tcp://127.0.0.1:2345") as client:
    future = client.submit(square, 4)
    print(future.result())  # 16

Client.submit() returns a standard Python future.

Graph computations

Scaler also supports graph tasks, for example:

from scaler import Client


def inc(i):
    return i + 1


def add(a, b):
    return a + b


def minus(a, b):
    return a - b


graph = {
    "a": 2,
    "b": 2,
    "c": (inc, "a"),  # c = a + 1 = 2 + 1 = 3
    "d": (add, "a", "b"),  # d = a + b = 2 + 2 = 4
    "e": (minus, "d", "c")  # e = d - c = 4 - 3 = 1
}

with Client(address="tcp://127.0.0.1:2345") as client:
    result = client.get(graph, keys=["e"])
    print(result)  # {"e": 1}

Nested computations

Scaler allows tasks to submit new tasks while being executed. Scaler also supports recursive task calls.

from scaler import Client


def fibonacci(clnt: Client, n: int):
    if n == 0:
        return 0
    elif n == 1:
        return 1
    else:
        a = clnt.submit(fibonacci, clnt, n - 1)
        b = clnt.submit(fibonacci, clnt, n - 2)
        return a.result() + b.result()


with Client(address="tcp://127.0.0.1:2345") as client:
    result = client.submit(fibonacci, client, 8).result()
    print(result)  # 21

IBM Spectrum Symphony integration

A Scaler scheduler can interface with IBM Spectrum Symphony to provide distributed computing across Symphony clusters.

$ scaler_symphony_cluster tcp://127.0.0.1:2345 ScalerService --base-concurrency 4

This will start a Scaler worker that connects to the Scaler scheduler at tcp://127.0.0.1:2345 and uses the Symphony service ScalerService to submit tasks.

Symphony service

A service must be deployed in Symphony to handle the task submission.

Here is an example of a service that can be used

class Message(soamapi.Message):
    def __init__(self, payload: bytes = b""):
        self.__payload = payload

    def set_payload(self, payload: bytes):
        self.__payload = payload

    def get_payload(self) -> bytes:
        return self.__payload

    def on_serialize(self, stream):
        payload_array = array.array("b", self.get_payload())
        stream.write_byte_array(payload_array, 0, len(payload_array))

    def on_deserialize(self, stream):
        self.set_payload(stream.read_byte_array("b"))

class ServiceContainer(soamapi.ServiceContainer):
    def on_create_service(self, service_context):
        return

    def on_session_enter(self, session_context):
        return

    def on_invoke(self, task_context):
        input_message = Message()
        task_context.populate_task_input(input_message)

        fn, *args = cloudpickle.loads(input_message.get_payload())
        output_payload = cloudpickle.dumps(fn(*args))

        output_message = Message(output_payload)
        task_context.set_task_output(output_message)

    def on_session_leave(self):
        return

    def on_destroy_service(self):
        return

Nested tasks

Nested task originating from Symphony workers must be able to reach the Scaler scheduler. This might require modifications to the network configuration.

Nested tasks can also have unpredictable resource usage and runtimes, which can cause Symphony to prematurely kill tasks. It is recommended to be conservative when provisioning resources and limits, and monitor the cluster status closely for any abnormalities.

Base concurrency

Base concurrency is the maximum number of unnested tasks that can be executed concurrently. It is possible to surpass this limit by submitting nested tasks which carry a higher priority. Important: If your workload contains nested tasks the base concurrency should be set to a value less to the number of cores available on the Symphony worker or else deadlocks may occur.

A good heuristic for setting the base concurrency is to use the following formula:

base_concurrency = number_of_cores - deepest_nesting_level

where deepest_nesting_level is the deepest nesting level a task has in your workload. If you have a workload that has a base task that calls a nested task that calls another nested task, the deepest nesting level is 2.

Performance

uvloop

For better async performance, you can install uvloop (pip install uvloop) and supply uvloop for the CLI argument --event-loop or as a keyword argument for event_loop in Python code when initializing the scheduler.

scaler_scheduler --event-loop uvloop tcp://127.0.0.1:2345

from scaler import SchedulerClusterCombo

scheduler = SchedulerClusterCombo(address="tcp://127.0.0.1:2345", event_loop="uvloop", n_workers=4)

Monitoring

From the CLI

Use scaler_top to connect to the scheduler's monitor address (printed by the scheduler on startup) to see diagnostics/metrics information about the scheduler and its workers.

$ scaler_top ipc:///tmp/127.0.0.1_2345_monitor

It will look similar to top, but provides information about the current Scaler setup:

scheduler          | task_manager         |   scheduler_sent         | scheduler_received
      cpu     0.0% |   unassigned       0 |   ObjectResponse      24 |          Heartbeat 183,109
      rss 37.1 MiB |      running       0 |         TaskEcho 200,000 |    ObjectRequest      24
                   |      success 200,000 |             Task 200,000 |               Task 200,000
                   |       failed       0 |       TaskResult 200,000 |         TaskResult 200,000
                   |     canceled       0 |   BalanceRequest       4 |    BalanceResponse       4
--------------------------------------------------------------------------------------------------
Shortcuts: worker[n] cpu[c] rss[m] free[f] working[w] queued[q]

Total 10 worker(s)
                 worker agt_cpu agt_rss [cpu]   rss free sent queued | object_id_to_tasks
W|Linux|15940|3c9409c0+    0.0%   32.7m  0.0% 28.4m 1000    0      0 |
W|Linux|15946|d6450641+    0.0%   30.7m  0.0% 28.2m 1000    0      0 |
W|Linux|15942|3ed56e89+    0.0%   34.8m  0.0% 30.4m 1000    0      0 |
W|Linux|15944|6e7d5b99+    0.0%   30.8m  0.0% 28.2m 1000    0      0 |
W|Linux|15945|33106447+    0.0%   31.1m  0.0% 28.1m 1000    0      0 |
W|Linux|15937|b031ce9a+    0.0%   31.0m  0.0% 30.3m 1000    0      0 |
W|Linux|15941|c4dcc2f3+    0.0%   30.5m  0.0% 28.2m 1000    0      0 |
W|Linux|15939|e1ab4340+    0.0%   31.0m  0.0% 28.1m 1000    0      0 |
W|Linux|15938|ed582770+    0.0%   31.1m  0.0% 28.1m 1000    0      0 |
W|Linux|15943|a7fe8b5e+    0.0%   30.7m  0.0% 28.3m 1000    0      0 |

• scheduler section shows scheduler resource usage
• task_manager section shows count for each task status
• scheduler_sent section shows count for each type of messages scheduler sent
• scheduler_received section shows count for each type of messages scheduler received
• function_id_to_tasks section shows task count for each function used
• worker section shows worker details, you can use shortcuts to sort by columns, the char * on column header show which column is sorted right now
  • agt_cpu/agt_rss means cpu/memory usage of worker agent
  • cpu/rss means cpu/memory usage of worker
  • free means number of free task slots for this worker
  • sent means how many tasks scheduler sent to the worker
  • queued means how many tasks worker received and queued

From the web UI

scaler_ui provides a web monitoring interface for Scaler.

$ scaler_ui ipc:///tmp/127.0.0.1_2345_monitor --port 8081

This will open a web server on port 8081.

Contributing

Your contributions are at the core of making this a true open source project. Any contributions you make are greatly appreciated.

We welcome you to:

• Fix typos or touch up documentation
• Share your opinions on existing issues
• Help expand and improve our library by opening a new issue

Please review our community contribution guidelines and functional contribution guidelines to get started 👍.

Code of Conduct

We are committed to making open source an enjoyable and respectful experience for our community. See CODE_OF_CONDUCT for more information.

License

This project is distributed under the Apache-2.0 License. See LICENSE for more information.

Contact

If you have a query or require support with this project, raise an issue. Otherwise, reach out to [email protected].