On 2 Apr 2015, stefan@agner.ch wrote:
To improve performance we remember the current page in the buffer and avoid reading it twice. This implicit page cache increases complexity while does not increase performance in real world cases. This patch removes that feature.
As discussed in the other patchset... http://thread.gmane.org/gmane.comp.boot-loaders.u-boot/215802
...I did some performance measurements:
Time to "Starting kernel ..."
- without bad block scan & with UBIFS fastmap: 2.02s
- with bad block scan & with UBIFS fastmap: 3.99s
- without bad block scan & without UBIFS fastmap: 4.42s
- with bad block scan & without UBIFS fastmap: 6.38s
Without page cache (with this patch applied): Time to "Starting kernel ..."
- without bad block scan & with UBIFS fastmap: 2.02s
- with bad block scan & with UBIFS fastmap: 4.01s
- without bad block scan & without UBIFS fastmap: 4.41s
- with bad block scan & without UBIFS fastmap: 6.39s
[snip]
I also measured 'write performance' with the mtd_speedtest (performing similar patch to the Linux driver) and I see no difference. I think a write benchmark is more appropriate to test this functionality? While at least it seems that neither read nor write is affected by the simplification.
Fwiw, Bill Pringlemeir.