Minimalist Async Evaluation Framework for R.

mirai

ミライ

( 未来 )

Minimalist Async Evaluation Framework for R

Designed for simplicity, a ‘mirai’ evaluates an R expression asynchronously in a parallel process, locally or distributed over the network, with the result automatically available upon completion.

Modern networking and concurrency built on nanonext and NNG (Nanomsg Next Gen) ensures reliable and efficient scheduling, over fast inter-process communications or TCP/IP secured by TLS.

Advantages include being inherently queued thus handling many more tasks than available processes, no storage on the file system, support for otherwise non-exportable reference objects, an event-driven promises implementation, and built-in asynchronous parallel map.

Quick Start

Use mirai() to evaluate an expression asynchronously in a separate, clean R process.

The following mimics an expensive calculation that eventually returns a vector of random values.

library(mirai)

m <- mirai({Sys.sleep(n); rnorm(n, mean)}, n = 5L, mean = 7)

The mirai expression is evaluated in another process and hence must be self-contained, not referring to variables that do not already exist there. Above, the variables n and mean are passed as part of the mirai() call.

A ‘mirai’ object is returned immediately - creating a mirai never blocks the session.

Whilst the async operation is ongoing, attempting to access a mirai’s data yields an ‘unresolved’ logical NA.

m
#> < mirai [] >
m$data
#> 'unresolved' logi NA

To check whether a mirai remains unresolved (yet to complete):

unresolved(m)
#> [1] TRUE

To wait for and collect the return value, use the mirai’s [] method:

m[]
#> [1] 5.735529 7.862045 6.024613 7.572171 5.791506

As a mirai represents an async operation, it is never necessary to wait for it. Other code can continue to be run. Once it completes, the return value automatically becomes available at $data.

while (unresolved(m)) {
  # do work here that does not depend on 'm'
}
m
#> < mirai [$data] >
m$data
#> [1] 5.735529 7.862045 6.024613 7.572171 5.791506

Daemons

Daemons are persistent background processes for receiving mirai requests, and are created as easily as:

daemons(4)
#> [1] 4

Daemons may also be deployed remotely for distributed computing and launchers can start daemons across the network via (tunnelled) SSH or a cluster resource manager.

Secure TLS connections can be used for remote daemon connections, with zero configuration required.

Async Parallel Map

mirai_map() maps a function over a list or vector, with each element processed in a separate parallel process. It also performs multiple map over 2D lists/vectors, allowing advanced patterns such as map over the rows of a dataframe or matrix.

df <- data.frame(
  fruit = c("melon", "grapes", "coconut"),
  price = c(3L, 5L, 2L)
)
m <- mirai_map(df, sprintf, .args = list(fmt = "%s: $%d"))

A ‘mirai_map’ object is returned immediately. Other code can continue to run at this point. Its value may be retrieved at any time using its [] method to return a list, just like purrr::map() or base::lapply(). The [] method also provides options for flatmap, early stopping and/or progress indicators.

m
#> < mirai map [3/3] >
m[.flat]
#> [1] "melon: $3"   "grapes: $5"  "coconut: $2"

All errors are returned as ‘errorValues’, facilitating recovery from partial failure. There are further advantages over alternative map implementations.

Design Concepts

mirai is designed from the ground up to provide a production-grade experience.

• Fast
  • Over 100x more responsive than common alternatives ^[1]
  • Built for low-latency applications such as real time inference or Shiny apps
• Reliable
  • Consistent behaviour with no reliance on global options or variables
  • Each mirai call is evaluated explicitly for transparent and predictable results
• Scalable
  • Launch millions of tasks simultaneously over thousands of connections
  • Proven track record handling heavy-duty workloads in the life sciences industry

mirai パッケージを試してみたところ、かなり速くて驚きました

Integrations

The following core integrations are documented, with usage examples in the linked vignettes:

Provides an alternative communications backend for R, implementing a new parallel cluster type, a feature request by R-Core at R Project Sprint 2023. ‘miraiCluster’ may also be used with foreach via doParallel.

Implements the next generation of completely event-driven, non-polling promises. ‘mirai’ may be used interchageably with ‘promises’, including with the promise pipe %...>%.

Asynchronous parallel / distributed backend, supporting the next level of responsiveness and scalability for Shiny. Launches ExtendedTasks, or plugs directly into the reactive framework for advanced uses.

Asynchronous parallel / distributed backend, capable of scaling Plumber applications in production usage.

Allows queries using the Apache Arrow format to be handled seamlessly over ADBC database connections hosted in background processes.

Allows Torch tensors and complex objects such as models and optimizers to be used seamlessly across parallel processes.

Powering Crew and Targets High Performance Computing

Targets, a Make-like pipeline tool for statistics and data science, has integrated and adopted crew as its default high-performance computing backend.

Crew is a distributed worker-launcher extending mirai to different distributed computing platforms, from traditional clusters to cloud services.

crew.cluster enables mirai-based workflows on traditional high-performance computing clusters using LFS, PBS/TORQUE, SGE and Slurm.

crew.aws.batch extends mirai to cloud computing using AWS Batch.

Thanks

We would like to thank in particular:

Will Landau for being instrumental in shaping development of the package, from initiating the original request for persistent daemons, through to orchestrating robustness testing for the high performance computing requirements of crew and targets.

Joe Cheng for optimising the promises method to work seamlessly within Shiny, and prototyping event-driven promises, which is implemented across nanonext and mirai.

Luke Tierney of R Core, for discussion on L’Ecuyer-CMRG streams to ensure statistical independence in parallel processing, and making it possible for mirai to be the first ‘alternative communications backend for R’.

Henrik Bengtsson for valuable insights leading to the interface accepting broader usage patterns.

Daniel Falbel for discussion around an efficient solution to serialization and transmission of torch tensors.

Kirill Müller for discussion on using ‘daemons’ to host Arrow database connections.

for funding work on the TLS implementation in nanonext, used to provide secure connections in mirai.

Installation

Install the latest release from CRAN or R-multiverse:

install.packages("mirai")

The current development version is available from R-universe:

install.packages("mirai", repos = "https://shikokuchuo.r-universe.dev")

Links & References

◈ mirai R package: <br / ◈ nanonext R package: https://shikokuchuo.net/nanonext/

mirai is listed in CRAN High Performance Computing Task View:
https://cran.r-project.org/view=HighPerformanceComputing

[1] Benchmark available in appendix of: https://shikokuchuo.net/user2024-conference/

–

Please note that this project is released with a Contributor Code of Conduct. By participating in this project you agree to abide by its terms.