KvikIO Reference — High-Performance GPU File IO

Review: 2026-10-01. Code below is illustrative unless explicitly described as CPU-tested. GPU execution, performance, GDS, and multi-GPU behavior require validation on target hardware.

KvikIO is a Python and C++ library for high-performance file IO. It provides bindings to NVIDIA cuFile, enabling GPUDirect Storage (GDS) — eligible local reads and writes between supported storage and GPU memory. When GDS isn't available, KvikIO falls back gracefully to POSIX IO while still handling both host and device data seamlessly.

KvikIO is part of the RAPIDS ecosystem and interoperates with CuPy, cuDF, Numba, and other GPU libraries.

Table of Contents

1. Installation
2. When to Use KvikIO
3. CuFile — Local File IO
4. RemoteFile — S3, HTTP, WebHDFS
5. Zarr Integration
6. Memory-Mapped Files
7. Runtime Settings
8. Performance Optimization
9. Interoperability
10. Common Patterns
11. Common Pitfalls

Installation

# CUDA 12.x
uv add "kvikio-cu12==26.8.*"

# CUDA 13.x
uv add "kvikio-cu13==26.8.*"

# For Zarr support (optional)
uv add "zarr==3.*"

Verify installation:

import kvikio
# Check if GDS is available
import kvikio.cufile_driver
print(kvikio.cufile_driver.get("is_gds_available"))  # True if GDS is set up

When to Use KvikIO

Use KvikIO when:

• Loading large binary data directly to GPU — avoids the CPU-memory copy that standard open() or NumPy's fromfile() would require
• Writing GPU arrays to disk — saves directly from device memory without copying to host first
• Reading from remote storage (S3, HTTP, WebHDFS) into GPU memory — convenient device destinations; libcurl uses host bounce buffers
• Working with Zarr arrays on GPU — the GDSStore backend reads chunks directly into CuPy arrays
• IO is the bottleneck — compare the actual storage, filesystem, access pattern and GDS eligibility
• Overlapping IO and compute — non-blocking reads/writes let you pipeline data loading with GPU computation

KvikIO is a poor fit when:

• Small transfers whose setup costs dominate measured latency
• You're reading structured formats (CSV, Parquet, JSON) — use cuDF instead, which has its own optimized readers
• You only need host memory — standard Python IO is simpler

CuFile — Local File IO

kvikio.CuFile is the primary interface for local file IO. It replaces Python's open() for GPU workloads.

Basic Usage

import cupy as cp
import kvikio

# Write a GPU array to disk
a = cp.arange(1_000_000, dtype=cp.float32)
with kvikio.CuFile("data.bin", "w") as f:
    f.write(a)

# Read it back
b = cp.empty(1_000_000, dtype=cp.float32)
with kvikio.CuFile("data.bin", "r") as f:
    f.read(b)

assert cp.all(a == b)

API Methods

File modes: "r" (read), "w" (write/truncate), "a" (append), "+" (read+write).

Non-Blocking IO with Futures

pread and pwrite split the operation into tasks executed in a thread pool and return an IOFuture:

import cupy as cp
import kvikio

data = cp.empty(10_000_000, dtype=cp.float32)

with kvikio.CuFile("data.bin", "r") as f:
    # Launch two non-blocking reads for different sections
    future1 = f.pread(data[:5_000_000])
    future2 = f.pread(data[5_000_000:], file_offset=5_000_000 * 4)

    # Do other work while IO happens...

    # Wait for completion
    bytes_read1 = future1.get()
    bytes_read2 = future2.get()

Partial Reads and Writes

import cupy as cp
import kvikio

# Read only a portion of a file
buf = cp.empty(1000, dtype=cp.float32)
with kvikio.CuFile("data.bin", "r") as f:
    # Read 1000 floats starting at byte offset 4000
    f.read(buf, size=4000, file_offset=4000)

Host Memory Support

KvikIO handles host memory transparently — no special API needed:

import numpy as np
import kvikio

# Write from host memory
a = np.arange(1_000_000, dtype=np.float32)
with kvikio.CuFile("data.bin", "w") as f:
    f.write(a)

# Read into host memory
b = np.empty_like(a)
with kvikio.CuFile("data.bin", "r") as f:
    f.read(b)

GDS Alignment

GDS works best with page-aligned IO. The relevant GDS I/O alignment is 4 KiB (4096 bytes); this is not a universal GPU memory-page size:

• File offset: should be a multiple of 4096
• Transfer size: should be a multiple of 4096

KvikIO handles unaligned IO correctly but splits it into aligned and unaligned parts, so aligned IO will be faster.

RemoteFile — S3, HTTP, WebHDFS

kvikio.RemoteFile reads remote files directly into GPU or host memory.

HTTP/HTTPS

import cupy as cp
import kvikio

buf = cp.empty(1_000_000, dtype=cp.float32)
with kvikio.RemoteFile.open_http("https://example.com/data.bin") as f:
    print(f.nbytes())  # File size
    f.read(buf)

AWS S3

import cupy as cp
import kvikio

# Using bucket + object name (requires AWS env vars or explicit credentials)
with kvikio.RemoteFile.open_s3("my-bucket", "data/file.bin") as f:
    buf = cp.empty(f.nbytes(), dtype=cp.uint8)
    f.read(buf)

# Using S3 URL
with kvikio.RemoteFile.open_s3_url("s3://my-bucket/data/file.bin") as f:
    buf = cp.empty(f.nbytes(), dtype=cp.uint8)
    f.read(buf)

# Public S3 (no credentials needed)
with kvikio.RemoteFile.open_s3_public("s3://public-bucket/data.bin") as f:
    buf = cp.empty(f.nbytes(), dtype=cp.uint8)
    f.read(buf)

# Presigned URL
with kvikio.RemoteFile.open_s3_presigned_url(presigned_url) as f:
    buf = cp.empty(f.nbytes(), dtype=cp.uint8)
    f.read(buf)

AWS credentials come from AWS_DEFAULT_REGION, AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, and optionally AWS_SESSION_TOKEN, or documented keyword arguments. Resolve temporary credentials explicitly; do not assume the complete boto3 profile/role provider chain. S3/public/presigned endpoints and HTTP range behavior must match the actual service.

Auto-Detect Endpoint Type

import kvikio

# KvikIO figures out the protocol from the URL
with kvikio.RemoteFile.open("s3://bucket/object") as f:
    ...

with kvikio.RemoteFile.open("https://example.com/file.bin") as f:
    ...

WebHDFS

import kvikio

with kvikio.RemoteFile.open_webhdfs("http://namenode:9870/webhdfs/v1/path/to/file") as f:
    buf = cp.empty(f.nbytes(), dtype=cp.uint8)
    f.read(buf)

Use the deployment's actual scheme, host, explicit port, and /webhdfs/v1/ path. This example does not configure Kerberos or establish access to a protected cluster.

Host Memory with RemoteFile

RemoteFile reads into host memory just as easily:

import numpy as np
import kvikio

with kvikio.RemoteFile.open_http("https://example.com/data.bin") as f:
    buf = np.empty(f.nbytes(), dtype=np.uint8)
    f.read(buf)

Zarr Integration

KvikIO provides a GPU store backend for Zarr (version 3.x). This enables reading and writing chunked N-dimensional arrays directly in GPU memory via GDS.

import zarr
from kvikio.zarr import GDSStore

# Enable GPU support in Zarr
zarr.config.enable_gpu()

# Create a GDS-backed store
store = GDSStore(root="data.zarr")

# Create and write a Zarr array (data stays on GPU)
z = zarr.create_array(
    store=store,
    shape=(1000, 1000),
    chunks=(100, 100),
    dtype="float32",
    overwrite=False,
)

# Reading returns CuPy arrays
chunk = z[:100, :100]  # Returns cupy.ndarray

Zarr + KvikIO is useful for:

• Climate/weather data (large multi-dimensional arrays)
• Bioinformatics (genomic arrays)
• Any workload using chunked arrays that need GPU processing

Requires a compatible Zarr 3 release in addition to KvikIO 26.08. Check codec GPU support; a GPU store alone does not make every codec device-resident.

Memory-Mapped Files

kvikio.mmap.Mmap provides memory-mapped file access with support for both host and device destinations:

from kvikio.mmap import Mmap
import cupy as cp

# Map a file for reading
with Mmap("data.bin", flags="r") as m:
    print(m.file_size())

    # Sequential read into device memory
    buf = cp.empty(1000, dtype=cp.float32)
    m.read(buf, size=4000, offset=0)

    # Parallel read (returns IOFuture)
    future = m.pread(buf, size=4000, offset=0)
    future.get()

Runtime Settings

KvikIO behavior is controlled via environment variables or the kvikio.defaults API.

Key Settings

Programmatic Configuration

import kvikio.defaults

# Query settings
print(kvikio.defaults.get("compat_mode"))
print(kvikio.defaults.get("num_threads"))

# Modify settings at runtime
kvikio.defaults.set({"num_threads": 16, "task_size": 8 * 1024 * 1024})

# Enable direct IO for reads
kvikio.defaults.set({"auto_direct_io_read": True})

Compatibility Mode

When GDS isn't available (missing libcufile.so, running in WSL, Docker without /run/udev), AUTO mode falls back to POSIX IO automatically. This fallback still requires a supported KvikIO/CUDA platform; it does not add macOS support or guarantee a speedup.

import kvikio.cufile_driver

# Check GDS capability; individual operations can still take compatibility/bounce-buffer paths
print(kvikio.cufile_driver.get("is_gds_available"))

cuFile Driver Configuration

import kvikio.cufile_driver

# Query driver properties
print(kvikio.cufile_driver.get("is_gds_available"))
print(kvikio.cufile_driver.get("major_version"))

# Configure settable properties
kvikio.cufile_driver.set("max_device_cache_size", 1024)

# Use as context manager (auto-reverts on exit)
with kvikio.cufile_driver.set({"poll_mode": True}):
    # poll mode active here
    ...
# poll mode reverted

Performance Optimization

1. Increase Thread Pool Size

The default of 1 thread is conservative. For large files, increase it:

import kvikio.defaults
kvikio.defaults.set({"num_threads": 16})

2. Use Non-Blocking IO for Pipelining

Overlap IO with compute by using pread/pwrite:

from pathlib import Path
import cupy as cp
import kvikio

def process_float32_chunks(path, chunk_items, consume):
    """Consume every chunk, including the tail, without reusing a live GPU buffer."""
    if chunk_items <= 0:
        raise ValueError("chunk_items must be positive")
    nbytes = Path(path).stat().st_size
    if nbytes % 4:
        raise ValueError("Input must contain complete float32 values")
    total = nbytes // 4
    if total == 0:
        return
    buf_a = cp.empty(chunk_items, dtype=cp.float32)
    buf_b = cp.empty_like(buf_a)
    stream = cp.cuda.get_current_stream()
    with kvikio.CuFile(path, "r") as f:
        count = min(chunk_items, total)
        future = f.pread(buf_a[:count], file_offset=0)
        offset = 0
        while count:
            if future.get() != count * 4:
                raise EOFError("Short read")
            next_count = min(chunk_items, total - offset - count)
            next_future = None
            if next_count:
                next_future = f.pread(buf_b[:next_count],
                                      file_offset=(offset + count) * 4)
            try:
                try:
                    consume(buf_a[:count])  # Enqueue only on the current CuPy stream.
                finally:
                    stream.synchronize()  # Finish using this buffer before recycling it.
            except BaseException:
                if next_future is not None:
                    next_future.get()  # Drain pending I/O before either buffer is freed.
                raise
            offset += count
            count, future = next_count, next_future
            buf_a, buf_b = buf_b, buf_a

# consume must finish using its input on this stream and must not retain that view.
# process_float32_chunks("large_data.bin", 1_000_000, consume)

The chunk/tail and buffer-lifetime logic is CPU-tested with fake asynchronous I/O and streams; actual overlap and throughput are not GPU-verified. A callback using other streams must join those streams before returning.

3. Align IO to Page Boundaries

GDS performs best with 4 KiB-aligned offsets and sizes:

# Good: aligned offset and size
f.read(buf, size=4096 * 1000, file_offset=4096 * 10)

# Slower: unaligned (KvikIO handles it, but splits into aligned + unaligned parts)
f.read(buf, size=5000, file_offset=100)

4. Enable Direct IO

For sequential writes and cold reads, Direct IO (bypassing OS page cache) can help:

import kvikio.defaults
kvikio.defaults.set({
    "auto_direct_io_read": True,
    "auto_direct_io_write": True,
})

5. Tune Task and Bounce Buffer Sizes

For very large files, increase task and bounce buffer sizes:

import kvikio.defaults
kvikio.defaults.set({
    "task_size": 16 * 1024 * 1024,       # 16 MiB per task
    "bounce_buffer_size": 64 * 1024 * 1024,  # 64 MiB bounce buffer
})

6. Page Cache Utilities

For benchmarking, clear the page cache to measure cold-read performance:

import kvikio

# Check page cache residency
pages_cached, total_pages = kvikio.get_page_cache_info("data.bin")
print(f"{pages_cached}/{total_pages} pages in cache")

# Drop the page cache for a single file (no elevated privileges needed; added in 26.04)
kvikio.drop_file_page_cache("data.bin")

# Avoid system-wide cache eviction on a shared host. Use a dedicated benchmark environment.

kvikio.clear_page_cache() is deprecated since 26.04 — use drop_system_page_cache() (or the per-file drop_file_page_cache()) instead.

Interoperability

With CuPy

KvikIO reads directly into CuPy arrays — this is the most common usage:

import cupy as cp
import kvikio

data = cp.empty(1_000_000, dtype=cp.float64)
with kvikio.CuFile("data.bin", "r") as f:
    f.read(data)
# data is now a CuPy array, ready for GPU computation

With Numba CUDA

KvikIO works with any buffer supporting the CUDA Array Interface:

from numba import cuda
import kvikio

d_arr = cuda.device_array(1_000_000, dtype="float32")
with kvikio.CuFile("data.bin", "r") as f:
    f.read(d_arr)

With cuDF

For raw binary data that isn't in a tabular format, use KvikIO to load, then convert:

import cupy as cp
import cudf
import kvikio

# Load raw float array, wrap as cuDF Series
buf = cp.empty(1_000_000, dtype=cp.float32)
with kvikio.CuFile("signal.bin", "r") as f:
    f.read(buf)
signal = cudf.Series(buf)

For tabular formats (CSV, Parquet, JSON, ORC), use cuDF's own readers — they're optimized for those formats.

With NumPy (Host Memory)

KvikIO seamlessly handles host memory:

import numpy as np
import kvikio

arr = np.empty(1_000_000, dtype=np.float32)
with kvikio.CuFile("data.bin", "r") as f:
    f.read(arr)

Common Patterns

Save and Load GPU Model Checkpoints

This is a raw contiguous-array container, not a self-describing or atomic checkpoint format. Save an ordered manifest with each array's name, shape, dtype/byte order, and byte offset; load using that exact order, and validate total file length and checksums. Synchronize producers before writing. For recovery-safe checkpoints, publish only after the payload and manifest are both complete.

import cupy as cp
import kvikio

def save_checkpoint(arrays: dict[str, cp.ndarray], path: str):
    """Save multiple GPU arrays to a single file."""
    with kvikio.CuFile(path, "w") as f:
        offset = 0
        for arr in arrays.values():
            if not arr.flags.c_contiguous:
                raise ValueError("Checkpoint arrays must be C-contiguous")
            if f.write(arr, file_offset=offset) != arr.nbytes:
                raise IOError("Incomplete checkpoint write")
            offset += arr.nbytes

def load_checkpoint(shapes_dtypes: dict, path: str) -> dict[str, cp.ndarray]:
    """Load GPU arrays from a checkpoint file."""
    arrays = {}
    with kvikio.CuFile(path, "r") as f:
        offset = 0
        for name, (shape, dtype) in shapes_dtypes.items():
            arr = cp.empty(shape, dtype=dtype)
            if f.read(arr, file_offset=offset) != arr.nbytes:
                raise IOError("Incomplete checkpoint read")
            offset += arr.nbytes
            arrays[name] = arr
    return arrays

Stream Data from S3 into GPU for Processing

import cupy as cp
import kvikio

with kvikio.RemoteFile.open_s3("my-bucket", "large-dataset.bin") as f:
    total_bytes = f.nbytes()
    if total_bytes % 4:
        raise ValueError("Object is not an integral number of float32 values")
    chunk_size = 100 * 1024 * 1024  # 100 MB chunks
    buf = cp.empty(chunk_size // 4, dtype=cp.float32)

    for offset in range(0, total_bytes, chunk_size):
        size = min(chunk_size, total_bytes - offset)
        bytes_read = f.read(buf[:size // 4], size=size, file_offset=offset)
        if bytes_read != size:
            raise EOFError("Short remote read")
        # Process chunk on GPU
        result = cp.mean(buf[:size // 4])
        cp.cuda.get_current_stream().synchronize()  # Finish before the next read reuses buf.

Replace Python open() for GPU Workloads

# Before: CPU-bound file IO
import numpy as np
data = np.fromfile("data.bin", dtype=np.float32)
import cupy as cp
gpu_data = cp.asarray(data)  # Extra copy: disk → CPU → GPU

# After: Direct to GPU
import cupy as cp
import kvikio
gpu_data = cp.empty(1_000_000, dtype=cp.float32)
with kvikio.CuFile("data.bin", "r") as f:
    f.read(gpu_data)  # disk → GPU directly (with GDS)

Common Pitfalls

1. Forgetting to set thread pool size — The default is 1 thread. For large files, kvikio.defaults.set({"num_threads": 16}) can dramatically improve throughput.

1. Using KvikIO for structured formats — Don't use KvikIO to read CSV/Parquet/JSON. Use cudf.read_csv(), cudf.read_parquet(), etc. KvikIO is for raw binary data.

1. Not checking GDS availability — Code works fine without GDS (falls back to POSIX), but won't get the full bandwidth benefit. Check with kvikio.cufile_driver.get("is_gds_available").

1. Misaligned IO in performance-critical paths — Use 4 KiB-aligned offsets and sizes for best GDS performance.

1. Not using context managers — Always use with kvikio.CuFile(...) to ensure files are properly closed and deregistered.

1. Expecting RemoteFile writes — RemoteFile is read-only. To write to remote storage, write locally first, then upload via the appropriate SDK (boto3 for S3, etc.).

1. Docker without GDS setup — In Docker, mount /run/udev read-only (--volume /run/udev:/run/udev:ro) for GDS to work. Otherwise, KvikIO silently falls back to POSIX.