Re: [lvs-users] lvs - how to estimate weights

To: " users mailing list." <lvs-users@xxxxxxxxxxxxxxxxxxxxxx>
Subject: Re: [lvs-users] lvs - how to estimate weights
From: "Thiago Varela" <thiago@xxxxxxxxxxxxxxxx>
Date: Fri, 14 Nov 2008 12:10:01 -0200
> so we can simply say that the weight can not be determined apriori using a
> math formula (before the momment we will start to use balancers)? I would
> like to have some aproximations in order to set up some weight before to
> start our cluster. after that, we can use tuning, to obtain an equilibrated
> distribution to all our servers (as malcolm said no more then 40%CPU usage
> on
> each).
> For example, if i have one P4 at 1.8Ghz and one P3 at 1.8Ghz, can i
> aproximate
> that P4 is twice good then P3 and set weight=2 on P4?
> Or, If we have 2 P4 servers at 1.8Ghz having 125M and 512M RAM, can we
> conclude that second server is 4 time better then first and set weight=4?
> Or, If we have two P4 servers at 2Ghz with the same amount of RAM, but on
> first server we are activating Hyper Threading technology, is correct to
> say
> that server with HT enabled is twice good then another and we can set
> weight=2 on it?

The answer to all this questions is NO.

Firstly because, as it has already been said, it is NOT POSSIBLE to
determine the influence of a given attribute (new instruction set, more RAM,
RAM speed, HT) over applications in general, as some will benefit from
certain attributes that another will not. It's hard to determine even if you
are talking about a specific application, as the usage type differs greatly
from server to server and thus benefitting more or less from different
attributes of the machine.

And secondly, in this proportions you mentioned, you are overestimative
about the real performance benefits of that hardware improvements. No P4
'willamette' (the only 1.8Ghz, AFAIK), with similar RAM, disk and
motherboard, will perform twice as fast as a P3 of the same frequency, in
practice. (Almost) No application will perform 4x faster in a server with
512MB RAM, than in an identical server with 128MB (well, unless your
application really NEEDS to use 512MB, but in this case you have a very very
bad dimensioning of your server, as the machine will just compulsively swap
all the time). The same goes for HT. If you had mentioned a sigle-core CPU
vs a dual-core one, then maybe, I said maybe, some applications could run
twice as fast. But HT? Even if your application benetifs very much from
threading, it will be very hard to see it  doubling it's performance just
from switching HT on.

<Prev in Thread] Current Thread [Next in Thread>