Use fs_read instead of fread

[lerabot]

I was doing some testing lately after resintalling ym BBA dev setup and felt like the BBA transfer speed was much slower than I remembered.

I stumbled on this post by TapamN and this help the solution:

One issue you might run into is slow file access over ethernet. Using the C library stdio.h functions (fread, fwrite) can be much slower than using the KOS filesystem calls directly (fs_read, fs_write) when reading/writing large blocks. With stdio, you get something like tens of KB/sec, while with KOS you can get over 1 MB/sec. Stdio might be faster when preforming many very small operations. dcload-serial doesn't have this issue.

There are the timing for loading a 256x256 .dtex texture (131kb)

FREAD
BBA: 3300ms
Serial: 6100ms

FS_OPEN
BBA: 70ms
Serial: 600ms

I'll test with GDEmu and CD and update my result later.

[TapamN]

So, the reason BBA is so slow is because newlib breaks up every I/O operation into tiny chunks, and dcload-ip is kind of half baked and doesn't handle many small requests well.

I feel incredibly stupid for not thinking about this earlier, but you can use setvbuf to disable stdio's buffering on a stream, which causes newlib to use large, single calls to KOS's fs_read/write operations, getting rid of the BBA's per-request overhead.

You would do something like this:

Code: Select all

FILE *f = fopen("/pc/blahblah", "r");
if (f) {
        setvbuf(f, NULL, _IONBF, 0);  //This makes it fast
        int amtread = fread(f, 1, sizeof(buff), buff);
}

[BB Hood]

Oh Snap I have to try that out!

[GyroVorbis]

Thanks for this... currently digging into what we can do about making Newlib's default buffering behavior less... stupid. lol.