[OpenAFS] 1.3.x memcache performance

Horst Birthelmer horst@riback.net
Mon, 27 Dec 2004 17:52:21 +0100


On Dec 27, 2004, at 5:39 PM, Kris Van Hees wrote:

> On Mon, Dec 27, 2004 at 05:28:14PM +0100, Andrej Filipcic wrote:
>>>> On gigabit network, the copy speed with memcache is never larger 
>>>> than
>>>> 20
>>>> mbyte/s, with disk cache it can go up to 70 mbyte/s.
>>
>> Well, it is not actually performance. dd if=/dev/zero to afs space 
>> transfer
>> reaches something like that. This is of course a peak number. Half of 
>> time
>> data is written to a local disk cache with no network activity, so on
>> average, the transfer would be 30-40 MB/s. AFS read speed is between 
>> 10-20
>> MB/s.
>>
>> New memcache parameters give similar write transfer now (~60 MB/s on 
>> average).
>
> So basically, you are measuring a combination of how fast you can 
> write to
> your cache, together with how fast the AFS client can flush the data 
> from
> the cache to the server.  Do you have writebehind turned on or off on 
> your
> AFS client?  If writebehind is allowed, then you only measure writing 
> to the
> cache + part of the flushing to the server (since flushbehind allows 
> the
> close() system call to return before the entire file has been flushed 
> to the
> server).

If writebehind is turned on (what I just don't assume) then what do you 
get??
The numbers won't be of any significance at all.

With the sync-on-close "on" you know how fast your client will be on a 
normal, unspectacular, daily usage.
IMHO that's the average speed you'll have to expect from that system 
afterwards in production.

Did I miss something?? Are we here in a client speed contest?? ;-)

Horst