[OpenAFS] Re: accessing R/O volume becomes slow

Hans-Werner Paulsen hans@MPA-Garching.MPG.DE
Fri, 28 Nov 2014 10:14:53 +0100

On 11/27/2014 01:11 PM, Stephan Wiesand wrote:
>> On 27 Nov 2014, at 11:26, Hans-Werner Paulsen <hans@MPA-Garching.MPG.DE> wrote:
>>> Yesterday, on another machine I created and deleted 4 million files on AFS. The number of afs_inode_cache slabs grew from 1 million to 5 million. Today there are still 5 million entries.
> It should shrink when there's memory pressure. If you're still worried, there's the -disable-dynamic-vcaches switch for afsd.
On my desktop PC (Linux 3.16.5 x86_64, OpenAFS 1.6.10) I set the 
-disable-dynamic-vcaches option, the -stat option has a value of 65536. 
When I create 100,000 files, I see 100,000 more afs_inode_cache slab 
objects. But, the fileserver is seeing this option, there are only 65253 
nFEs, 65253 nCBs (4194304 nblks). Without -disable-dynamic-vcaches the 
number of CBs is about the number of created files. And if I try to 
create more files than nCBs on the fileserver, the fileserver 
(dafileserver) hangs for about 15 minutes (dafileserver 100-120% cpu!), 
and I get a "connection timeout" on the client.