Re: [PATCH 5/8] knfsd: repcache: dynamically expand under load

[Greg Banks <gnb@sgi.com>]

On Mon, Oct 16, 2006 at 12:14:01PM +1000, Neil Brown wrote:
> On Wednesday October 11, gnb@melbourne.sgi.com wrote:
> >  /*
> > + * Decide whether it is time to expand the cache.  Returns 1 iff
> > + * the cache is to be expanded.  Called with bucket lock held.
> > + */
> > +static int
> > +nfsd_cache_expand_ratelimit(struct svc_cache_bucket *b)
> > +{
> > +	unsigned long now = jiffies;
> > +
> > +	b->nhits++;
> > +	if (b->last == 0) {
> > +		b->last = now;
> > +	} else if ((now - b->last) > CACHE_RATE_JIFFIES &&
> > +		   b->nhits > (b->size >> 4)) {
> > +		b->nhits = 0;
> > +		b->last = now;
> > +		return 1;
> > +	}
> > +	return 0;
> > +}
> 
> I'm wondering a bit about this function.
> Firstly (and this is very minor) I expected to see "time_before" or
> similar when comparing times.  The code you have is correct, but maybe
> it will also be obvious if it read
>            if (time_after(now, b->last + CACHE_RATE_JIFFIES))
> ??

Ok.

> Also, nhits is only reset when we return 1, and hence grow the bucket.
> So if we only occasionally had large bursts which included
> CACHE_BUCKET_MAX_SIZE requests in CACHE_THRESH_AGE jiffies, we would
> still eventually grow the bucket to it's max size.
> I think I would like to have something like
> 
> 
> 	b->nhits++;
> 	if (b->last == 0) {
> 		b->last = now;
> 	} else if (time_after(now, b->last + CACHE_MAX_WAIT)) {
> 		b->last = now; /* too slow, forget it */
> 		b->nhits = 0;
> 		return 0;
> 	} else if (time_after(now, b->last + CACHE_RATE_JIFFIES) &&
> 		   b->nhits > (b->size >> 4)) {
> 		b->nhits = 0;
> 		b->last = now;
> 		return 1;
> 	}
> 	return 0;
> 
> with CACHE_MAX_WAIT maybe 60*HZ.
> 
> Make sense?
> 

How about this?

 	b->nhits++;
 	if (b->last == 0) {
 		b->last = now;
 	} else if (time_after(now, b->last + CACHE_MAX_WAIT)) {
 		b->last = now; /* too slow */
 		b->nhits >>= 1;
 		return 0;
 	} else if (time_after(now, b->last + CACHE_RATE_JIFFIES) &&
 		   b->nhits > (b->size >> 4)) {
 		b->nhits = 0;
 		b->last = now;
 		return 1;
 	}
 	return 0;
 
It gives a damping factor without totally forgetting the past hits.
Otherwise the case of a regular burst every 61 seconds fails to expand
the repcache, which I think it should.


Greg.
-- 
Greg Banks, R&D Software Engineer, SGI Australian Software Group.
I don't speak for SGI.

-------------------------------------------------------------------------
Using Tomcat but need to do more? Need to support web services, security?
Get stuff done quickly with pre-integrated technology to make your job easier
Download IBM WebSphere Application Server v.1.0.1 based on Apache Geronimo
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=120709&bid=263057&dat=121642
_______________________________________________
NFS maillist  -  NFS@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/nfs