perf(cuda): read pooled files in 8 MiB chunks (#10715)

## Problem

Splitting staged file reads into 4 MiB chunks adds per-chunk I/O and transfer overhead.

## Solution

Increase physical read and H2D chunks to 8 MiB, retaining concurrency 32 and existing logical batches.

## Assessment

Uses larger staging buffers per request. Independent of mapped-file copying; based on #10726.

Signed-off-by: Alexander Droste <alexander.droste@protonmail.com>

Author: 0ax1

vortex-cuda/README.md

@@ -10,6 +10,11 @@ Key files:
 - `vortex-cuda/src/arrow/canonical.rs`: canonical-array export to `ArrowDeviceArray`.
 - `vortex-test/e2e-cuda/src/lib.rs`: cuDF interop harness.
 
+## CUDA file scans
+
+Local pooled file reads use fixed 8 MiB chunks with concurrency 32. These are physical
+file-read/H2D chunks, independent of logical row ranges or scan batch sizes.
+
 ## Building cuDF for Arrow Device interop
 
 The `cudf-test-harness` repository provides prebuilt cuDF binaries for Arrow Device interop testing on

vortex-cuda/ffi/README.md

@@ -24,6 +24,9 @@ Use `vx_cuda_scan_path_arrow_device_stream` to read such a local file through pi
 and receive an Arrow C Device stream. Reuse the same CUDA session across scans so the pinned buffer
 pool and CUDA state are reused as well.
 
+Pooled data reads use fixed 8 MiB chunks and concurrency 32;
+scan row ranges and batch sizes are unchanged.
+
 On Linux, use `vx_cuda_scan_path_arrow_device_stream_with_options` with
 `vx_cuda_scan_options.flags = VX_CUDA_SCAN_FLAG_DIRECT_IO` to bypass the operating system page
 cache for pooled data-plane reads. Footer and zone-map reads remain buffered on the host.

vortex-cuda/src/pooled_read_at/file.rs

@@ -38,7 +38,7 @@ pub const DEFAULT_FILE_CONCURRENCY: usize = 32;
 
 // Physical segments can exceed the coalescing limit. Split their I/O, not their encoding,
 // to limit blocking read sizes and start transfers before the whole segment is read.
-const FILE_READ_CHUNK_BYTES: usize = 4 << 20;
+const FILE_READ_CHUNK_BYTES: usize = 8 << 20;
 
 /// Options controlling how [`PooledFileReadAt`] opens and reads a local file.
 #[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]

vortex-cuda/src/pooled_read_at/file/tests.rs

@@ -119,7 +119,7 @@ impl FileReadBackend for FakeFileReadBackend {
 }
 
 fn file_byte(offset: u64) -> u8 {
-    // Unlike a repeating u8 counter, this distinguishes adjacent 4 MiB chunks.
+    // Unlike a repeating u8 counter, this distinguishes adjacent read chunks.
     (offset % 251) as u8
 }