Chow Chi Ho (FP&HO)
chchow at cabletv.com.hk
Wed May 2 04:09:06 UTC 2007
how to define "all at once" ?
From: dhcp-users-bounce at isc.org [mailto:dhcp-users-bounce at isc.org]On Behalf Of Simon Hobson
Sent: Wednesday, May 02, 2007 6:41 AM
To: dhcp-users at isc.org
Subject: Re: dhcpd performance?
Jiann-Ming Su wrote:
> > define "all at once." (a reoccurring question that you will see
>in the archives)
>> "when 1000 people show up for work" is much different than "wake
>>on lan" at 8am.
>> /me thinks of 1000 voip streams. The dhcp server will not be the
>> what you got, get the system up, fix the problems.
>This is what I'm hearing... upgrading the master phone switch will
>cause all the phones hanging off said switch to reboot. Yes, it would
>make sense to NOT upgrade all the switches at once. But, I can't
>control what the phone folks do and don't do.
Which switch are you referring to, the ethernet switches to which the
phones are connected and which I assume are powering the phones, or
the call processing system that's running it ?
If it's the PoE switches then I'd say it would be hard to reset all
of them at once - short of having a 'big red switch'. Even then, I
suspect that the switches will actually power up the ports in
If it's the call processor, then that implies that it is sending a
message to each device telling it to reboot. I think you are now
introducing a small spread to the reboot times of individual phones -
whilst it could use broadcasts, I suspect it would have to send out a
packet to each device so as to handle the general case where phones
could be anywhere.
This is largely hypothetical anyway. If the dhcp server could only
manage 1 client/s then that's under 20 minutes, 10/s cuts that to
under 2 minutes to service all 1000 devices - that should be no
problem for ANY reasonable client which will back off and retry for
The post David referenced earlier suggested 40
discover/offer/request/ack cycles/second for a moderate system - that
would get your 'boot 1000 clients' time to something in the order of
25 seconds. In reality, many clients will have retried in that time,
so perhaps double that - but you are still talking of only a minute
assuming that you boot all 1000 clients in that short a timescale.
If you think that won't be acceptable for a rare event, then I think
you need to investigate further into EXACTLY what sort of thing may
go on. Things like, how fast do the ethernet switches power up the
client ports after a power interruption, how fast does the call
processor get the devices to reboot, how soon and how often to
clients retry their DHCP requests, do the clients keep their address
(so they can just do a request/ack) or do they always cold start.
More information about the dhcp-users