Load times for Large Subnets

[David W. Hankins]

On Tue, Oct 09, 2007 at 04:35:02PM -0700, Larry Apolonio wrote:
> Oct  9 16:32:11 dhcp-amd dhcpd: Lease IP hash:  Contents/Size (%): 
> 410524/100003 (410%). Min/max: 2/7
> Oct  9 16:32:11 dhcp-amd dhcpd: Lease UID hash: Contents/Size (%): 
> 200000/100003 (199%). Min/max: 0/101
> Oct  9 16:32:11 dhcp-amd dhcpd: Lease HW hash:  Contents/Size (%): 
> 200000/100003 (199%). Min/max: 0/5

vs:

> Oct  9 15:40:25 dhcp-amd dhcpd: Lease IP hash:  Contents/Size (%):
> 410524/20000003 (2%). Min/max: 0/1
> Oct  9 15:40:26 dhcp-amd dhcpd: Lease UID hash: Contents/Size (%):
> 200000/20000003 (0%). Min/max: 0/101
> Oct  9 15:40:26 dhcp-amd dhcpd: Lease HW hash:  Contents/Size (%):
> 200000/20000003 (0%). Min/max: 0/1

It's curious that the UID hash turns out the same in both tables...and
so horribly wrong; max of 101 entries in one bucket!

It's hard to tell if it's going more wrong in the smaller table size.
I should have put in mean/std.dev as well.


I wouldn't have suspected such a big difference in performance for
such a (evidently) small change in hash distribution.  I'd literally
expect to see more like '0/137<->0/1' than '0/7<->0/1'  Possibly mean
and std.dev would give us a different idea of how much things have
shifted...I should have put those in the reporter but I was busy
looking at a different corner of the problem.

Could you mail your dhcpd.conf and dhcpd.leases to dhcp-bugs at isc.org
(or me gpg encrypted)?  When we have time, it looks like we need to
look more closely at our "opaque string of bits" hash function.  0/5
is really quite bad at 200% load.


> What do those IP hash %ages mean?

That's how loaded the table is.  Just doing the math for the
contents/size so I didn't have to do it in my head.

At 410%, I would have expected to see a min/max of 4/5 for perfect
distribution.  2/7 isn't so far away from that...nothing that would
explain the ludicrous shift in performance.  The "0" in the uid/hw
hashes is pretty telling; some buckets aren't getting any leases at
all.

-- 
Ash bugud-gul durbatuluk agh burzum-ishi krimpatul.
Why settle for the lesser evil?	 https://secure.isc.org/store/t-shirt/
-- 
David W. Hankins	"If you don't do it right the first time,
Software Engineer		     you'll just have to do it again."
Internet Systems Consortium, Inc.		-- Jack T. Hankins