[Lustre-discuss] How to you generated discarded readahead pages?
Tom.Wang
Tom.Wang at Sun.COM
Mon Mar 24 01:12:45 PDT 2008
Mark Seger wrote:
> I'm running a little script that does 3 consecutive small reads of 4K
> each and then sleeps for a second and as you can see from the following
> collectl output readahead is happening. How do I know? Because the
> network traffic is so high:
>
> # <----------Network----------><-------------Lustre
> Client-------------->
> #Time netKBi pkt-in netKBo pkt-out Reads KBRead Writes KBWrite
> Hits Misses
> 09:31:54 1084 740 29 394 3 12 0 0
> 0 3
> 09:31:55 1084 740 29 394 3 12 0 0
> 0 3
> 09:31:56 1084 740 29 394 3 12 0 0
> 0 3
> 09:31:57 1084 741 29 394 2 8 0 0
> 0 2
>
> I'm getting my data from
> /proc/fs/lustre/llite/spfs1-00000100768c2800/read_ahead_stats and when I
> look at the disards, they're not changing. What I would like to know is
> why. as I'd think all that extra data would be discarded from the
> cache. Is it because some other process needs to put pressure on it?
These read-ahead pages will be discarded when it met memory pressure or
the client revokes the lock of these pages.
> I
> have 4GB memory and just wrote a 5GB file which should certain force out
> anything already there but the counter never changed.
>
If you did not change the parameters in /proc, lustre client only permit
32M dirty cache currently,
you can check max_dirty_mb in /proc . So writing 5GB file will not
trigger memory pressure on client.
> I do see "readpage not consecutive" changing,
>
> -mark
>
>
> _______________________________________________
> Lustre-discuss mailing list
> Lustre-discuss at lists.lustre.org
> http://lists.lustre.org/mailman/listinfo/lustre-discuss
>
More information about the lustre-discuss
mailing list