Hi Chunhai,
On 2025/9/22 20:43, Chunhai Guo wrote:
Direct I/O is particularly useful in memory-sensitive scenarios, as it
provides more predictable performance by avoiding unnecessary page cache
overheads. For example, when accessing large files such as AI model files
that are typically read only once, buffered I/O introduces redundant page
cache usage and extra page copies, leading to unstable performance and
increased CPU load due to memory reclaim. While Direct I/O can avoid these.
The table below shows that the performance of direct I/O is up to 54.6%
higher than buffered I/O in the low-memory scenario. The results were
obtained using the fio benchmark with 8 threads, each thread read a 2.5GB
file, on ARM64 Android devices running the 6.6 kernel with an 8-core CPU
and 12GB of memory.
+--------------------------------------------------------------------------+
| fio benchmark | buffered I/O (MiB/s) | direct I/O (MiB/s) | diff |
|---------------------+----------------------+--------------------+--------|
| normal scenario | 2629.8 | 3648.7 | +38.7% |
|---------------------+----------------------+--------------------+--------|
| low memory scenario | 2350.0 | 3633.9 | +54.6% |
+--------------------------------------------------------------------------+
Thanks for your patch!
Yes, avoid page cache by using direct I/O for read-once data (e.g.
distributing huge LLM model) actually makes sense on my side and your test
result is impressive.
I will look into your implementation later, since it's too late for v6.18.
Let's address this feature for the v6.19 cycle.
Thanks,
Gao Xiang
This patch does not support the following two cases. They will fall back to
buffered I/O:
(1) large folios, which will be supported in a follow-up patch.
(2) folios with private data attached, as the private data is required by
this direct I/O implementation.
Signed-off-by: Chunhai Guo <[email protected]>
