Increase FILLUNIT (currently 5kB) in main/rfc1867.c otherwise large uploads are inefficients and slow

Description

I recently stumbled upon a very simple file upload code which was quite slow. It turned out that it was IO-limited while writing the temporary file to the disk. A strace showed that PHP would spoonfeed the incoming network data to the disk :

16:51:04 read(6, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 4142) = 4142
16:51:04 read(6, "\1\5\0\1 \0\0\0", 8)  = 8
16:51:04 read(6, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 977) = 977
16:51:04 write(8, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 5119) = 5119
16:51:04 read(6, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 5119) = 5119
16:51:04 write(8, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 5119) = 5119
16:51:04 read(6, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 2096) = 2096
16:51:04 read(6, "\1\5\0\1 \0\0\0", 8)  = 8
16:51:04 read(6, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 3023) = 3023
16:51:04 write(8, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 5119) = 5119

On the network/read side this is not very efficient, it uses a lot of syscalls (especially since there is a recurring 8-byte read() call). On my experiment (a n1-standard-8 GCE VM), a single process would use 50% CPU (on the system side, not the user one) with an upload proceeding at 15 MB/s.

On the disk/write side this is also inefficient, and runs quickly in real cloud limits : in my case a 100G "SSD" disk which GCE advertises as 48 MB/s and 3000 IO/s max; my benchmarks with fio confirmed those figures.

The problem with such a small FILLUNIT=5kB buffer, is that you can't get a better throughput than 3000*5 kB/s =15 MB/s. So you don't get the max througput of your disk, you top on IOPS first.

Of course that is amplified by the fact that is a VM and uses - as recommended by Google - the 'noop' IO scheduler and there is no write coalescing. I guess an IO scheduler could mitigate this on a bare-metal local storage. But I think running PHP on such a cloud VM (8 vCPUs/ 32GB RAM/ 100G "SSD") is very common. And the performance drop (and global pressure on the server storage) is very noticeable.

According to https://github.com/php/php-src/commit/44b68122c2ba0cf07cd837af160b3bd01527d081 FILLUNIT has been 5 kB for 23 years.

I would suggest it is time to bump it up. I would suggest at least 64 kB, GCE suggests 256 kB (https://cloud.google.com/compute/docs/disks/optimizing-pd-performance). Only one instance of this buffer is used at once, during the file upload loop (thus per-process), and can be released as soon as the loop is done.

Note that the value is used to allocate a second transient buffer on the stack in multipart_buffer_read_body() - and that might abuse the stack (which defaults to 1MB on Windows and 8MB on Linux if I'm not mistaken).

PHP Version

PHP 8.1.27

Operating System

Debian 10

从 main/rfc1867.c 开始，定位 FILLUNIT 和 multipart_buffer_read_body()，然后检查上传期间缓冲区大小是如何分配的。将报告中的 strace 和 fio 测量结果作为性能背景，并考虑临时栈缓冲区。当一个达成共识的更大值能够提高上传吞吐量，同时不影响上传行为或栈使用情况时，即表示完成。