On Wednesday 12 August 2009 16:00:40 Andrew McKay wrote:
> Hey Charles,
>
> When testing with the 2GB NAND I was visited by the OOM killer a few times.
> This makes me think we're short on RAM for handling a 2GB NAND part. Our
> board currently has 32MB of RAM, of which 8MB is used for a RAM disk. When
> I dropped the ramdisk down to 3.5MB for testing purposes, I didn't have
> issues with the OOM killer any more. We're looking at moving up to 64MB of
> RAM to avoid this issue. However in the future I'd like to be able to
> estimate the memory usage of YAFFS2 based on NAND size.
>
> I found a thread about YAFFS2 memory usage, and I just want to make sure I
> understand it correctly.
>
> http://www.yaffs.net/lurker/message/20090701.190059.23524635.ca.html#yaffs
>
> > * yaffs_Objects: Each object (file, directory,...) holds a yaffs_Object
>
> in memory which is around 120 bytes per object.
>
> So every file, directory, etc. uses up 120 bytes of RAM. This is all the
> time? Right from when the filesystem is mounted? So if I have 1000 objects
> on the device I'll be using up 120000 bytes?
>
> > * yaffs_Tnodes: These are the things used to build the trees.
>
> The part I'm using is 8192 erase blocks, and 64 pages per erase block.
> That means there are 524288 chunks in my filesystem. Using your equation I
> come up with
>
> Log2(524288) = 19 bits
> 19 + 1 = 20 (which is already even)
>
> So 20 bits will be used to represent each chunk.
>
> Assuming worst case and the filesystem is full I will be using all 524288
> chunks. This means that I will need 20 * 512K which is 10MB of RAM to
> store all the Tnodes. Does that seem about right?
Bits mate, not bytes... Therefore that calc should come out to be closer to
1.2Mbytes
The calculation is somewhat complicated by the fact that Tnodes are managed in
arrays of 16, ie 40 bytes in this case. Thus:
* On average larger files will have a wasted 20 bytes of Tnode.
* Very small files will still need 40 bytes of Tnode, even if they only use a
small amount of that.
Thus lots of small files will skew the numbers a bit.
If you do a /proc/yaffs you can get the actual numbers in use and use those
nTnodesCreated * 40 bytes
+
nObjectsCreated * approx 120 bytes
It would be possible to make a tweak to handle very short files better. Files
smaller than 1 chunk don't really need a tnode tree since the tnode pointer
could be stored directly in the object structure.
>
> I was also copying my Linux source tree to NAND. It's about 22527 files.
> This would mean it would require 120 * 22527, or about 2.6MB of RAM for all
> of the Objects.
>
> Of course as you mentioned in the email there is some other overhead on top
> of this, but this should be a large portion of the memory required to
> handle a YAFFS filesystem?
>
> Thanks again,
> Andrew McKay
> Iders Inc.
>
>
> _______________________________________________
> yaffs mailing list
> yaffs@lists.aleph1.co.uk
> http://lists.aleph1.co.uk/cgi-bin/mailman/listinfo/yaffs
(text/plain)