vendredi, décembre 31, 2010

libFAT : hide and fseek ...

Well, initially I thought that the performance issue of TCP transfers would be worth another investigation-novel-styled post, but being somewhere between Christmas and New Year's Eve didn't allowed me to do so. Or maybe just digesting the source code of libfat, figuring out what kind of access pattern made it so slow and what more subtle access pattern was making my new cache policy go wrong was already demanding enough.

The patch is not ready for inclusion in libfat according to devkitpro's standards. Yet, having a look at it is the only way for you to have an idea of what can be changed to fix the performance issue.

The core idea of this patch is to allow pages in the cache to reflect only partially what is (should be) on disk. The new readable field gives the position in the page where we leave 'valid' data and enter uncharted, here-be-dragons, area which content shouldn't be relied upon. I proceeded with additional extensive checks (64M movies transfers, md5 checksum verification and fsck on the media card) and fixed one last bug affecting end of files. The patch released on Tuesday is thus invalidated.

Remember, kids: FAT is the worst filesystem ever.

Oh, and btw, server.pl, the PC-side data source for a PC-to-DS wifi transfer can now be invoked so that it will cycle within a list of files, allowing one to efficitently transfer the files one after the other, e.g. :

server.pl neoflash/*.nds [DS_IP_ADDR [PC_IP_ADDR]]

Aucun commentaire: