[Ocfs2-devel] Performance issue about localalloc

[Ocfs2-devel] Performance issue about localalloc

[Joseph Qi]

Hi Mark,
I have encountered a performance issue about localalloc.

ocfs2 volume information:
block size = 4k, cluster size = 1M, mount option with localalloc=2048

So in the above circumstance, gd size is 32256M. Suppose I want to
create a normal file with size of 50G. Since the localalloc window is
2G, it will eventually allocate space like, 2Gfrom gd0, 2G from gd1,
..., and this will lead to poor performance because of the
dis-contiguous file with large span.

One way to resolve this issue is changing the mount option with
localalloc=31104, but this will make end user confused because 31G is
"missing" even he has only used 2M.

So I wander if we can let the allocation continues with the last gd if
it has enough space for a localalloc window. I have blamed the code
history and found that firstly you have already designed like this but
disabled in the commit "ocfs2: properly set and use inode group alloc
hint". I wander why?

A simply code hack is,
static u64 ocfs2_group_from_res(struct ocfs2_suballoc_result *res)
{
	if (res->sr_blkno == 0)
		return 0;  // This will set the hint to 0
	...
}

Am I missing something?

Thanks,
Joseph

[Ocfs2-devel] Performance issue about localalloc

[Joseph Qi]

Hi Mark,
Could you please take a look at this?
And I want to change it by setting hint to res->sr_bg_blkno if it is
cluster alloc.
From my test case (block size = 4K, cluster size = 1M, localalloc
2048M), the file allocation is more contiguous than before and has much
better performance.

 	status = ocfs2_search_chain(ac, handle, bits_wanted, min_bits,
 				    res, &bits_left);
 	if (!status) {
-		hint = ocfs2_group_from_res(res);
+		if (ocfs2_is_cluster_bitmap(ac->ac_inode))
+			hint = res->sr_bg_blkno;
+		else
+			hint = ocfs2_group_from_res(res);
 		goto set_hint;
 	}
 	if (status < 0 && status != -ENOSPC) {

On 2015/9/10 19:23, Joseph Qi wrote:
> Hi Mark,
> I have encountered a performance issue about localalloc.
> 
> ocfs2 volume information:
> block size = 4k, cluster size = 1M, mount option with localalloc=2048
> 
> So in the above circumstance, gd size is 32256M. Suppose I want to
> create a normal file with size of 50G. Since the localalloc window is
> 2G, it will eventually allocate space like, 2Gfrom gd0, 2G from gd1,
> ..., and this will lead to poor performance because of the
> dis-contiguous file with large span.
> 
> One way to resolve this issue is changing the mount option with
> localalloc=31104, but this will make end user confused because 31G is
> "missing" even he has only used 2M.
> 
> So I wander if we can let the allocation continues with the last gd if
> it has enough space for a localalloc window. I have blamed the code
> history and found that firstly you have already designed like this but
> disabled in the commit "ocfs2: properly set and use inode group alloc
> hint". I wander why?
> 
> A simply code hack is,
> static u64 ocfs2_group_from_res(struct ocfs2_suballoc_result *res)
> {
> 	if (res->sr_blkno == 0)
> 		return 0;  // This will set the hint to 0
> 	...
> }
> 
> Am I missing something?
> 
> Thanks,
> Joseph
> 
> 
> _______________________________________________
> Ocfs2-devel mailing list
> Ocfs2-devel at oss.oracle.com
> https://oss.oracle.com/mailman/listinfo/ocfs2-devel
> 
>