Zero-Copy Julia to R Array Bridge via ALTREP.

jlview — Zero-Copy Julia↔R Arrays via ALTREP

jlview provides zero-copy R views of Julia-owned arrays using R’s ALTREP (Alternative Representations) framework. Instead of copying data between Julia and R, jlview returns lightweight R vectors that point directly into Julia’s memory.

Benchmark: 50K × 25K Float64 matrix (9.3 GB)

Installation

# install.packages("remotes")
remotes::install_github("tanaylab/jlview")

Requirements

• R ≥ 4.0
• Julia ≥ 1.6
• JuliaCall R package

Usage

library(jlview)
JuliaCall::julia_setup()

# Create a Julia array
JuliaCall::julia_command("x = randn(10000, 1000)")

# Zero-copy view — R sees Julia's memory directly
m <- jlview(JuliaCall::julia_eval("x"))
dim(m) # [1] 10000  1000
sum(m) # works natively, no data copied

# Named arrays are supported
JuliaCall::julia_command("using NamedArrays")
JuliaCall::julia_command("named = NamedArray(randn(3), [\"a\", \"b\", \"c\"])")
v <- jlview_named_vector(JuliaCall::julia_eval("named"))
v["a"] # access by name, still zero-copy

# Sparse matrices (zero-copy values, shifted indices)
JuliaCall::julia_command("using SparseArrays")
JuliaCall::julia_command("sp = sprand(1000, 500, 0.01)")
s <- jlview_sparse(JuliaCall::julia_eval("sp"))
class(s) # "dgCMatrix"

# Explicit release to free Julia memory early
jlview_release(m)

Key Features

• Zero-copy dense arrays — Float64, Int32 map directly to R’s REALSXP, INTSXP
• Type conversion — Float32, Int64, Int16 are converted once in Julia, then zero-copy to R
• Named arrays — NamedArray row/column names attached atomically without triggering copy
• Sparse matrices — dgCMatrix with zero-copy values (nzval) and shifted indices
• Copy-on-write — R’s standard COW semantics: reads are zero-copy, writes trigger materialization
• GC safety — Julia arrays are pinned while R holds references; three-layer defense (pinning dict, memory pressure tracking, explicit release)
• Fork safety — Safe with parallel::mclapply() (PID-guarded finalizers)
• Serialization — saveRDS()/readRDS() work correctly (materializes on save)

Supported Types

How It Works

jlview uses R’s ALTREP framework to create R vectors backed by Julia memory:

1. Pin — The Julia array is stored in a global dictionary, preventing Julia’s GC from collecting it
2. Wrap — An ALTREP R vector is created whose Dataptr returns Julia’s raw data pointer
3. Use — R operations (sum, subsetting, etc.) read directly from Julia’s memory
4. Release — When R garbage-collects the ALTREP object, a C finalizer calls Julia to unpin the array

The entire pin→ALTREP→finalizer path is implemented in C using Julia’s C API (jl_call1), avoiding JuliaCall overhead and ensuring safety during R’s garbage collection.