[92] in Zephyr Mailing List
Server redesign
daemon@ATHENA.MIT.EDU (Richard Basch)
Thu Jul 8 20:30:43 1993
Date: Thu, 08 Jul 93 18:25:04
From: probe@MIT.EDU (Richard Basch)
To: jis@MIT.EDU, mar@MIT.EDU, tytso@MIT.EDU, shapere@MIT.EDU, wdc@MIT.EDU
Cc: zephyr@MIT.EDU
As you all know, Zephyr did not fare that well at the end of last
term. The servers had CPU utilizations of about 85%+ as it came
close to collapsing. This suggests that we will not be able to
solve the problem with more memory (virtual or real); the problem
is that the message delivery and subscription handling is too much
for the server.
I believe that some optimizations can be done with the current server
to gain some leeway into next term, but I believe we need to sit down
and redesign the server.
The current growth of Zephyr usage suggests that even these
optimizations will not last the academic year. A year ago, I predicted
the collapse of the servers with a doubling of usage, and as it stands
now, we have more than doubled, and it has exceeded the capacity of
3 maxines as opposed to the 2 VS-IIs when this prediction was made.
The growth in Zephyr traffic has been greater than the growth of Athena
logins (this is a subjective assessment), and with the onset of ResNet,
it will only grow faster. It is my hope that we can redesign the system
quickly and begin implementing the solution quickly. I am hoping that
ResNet will not take off for about a year, because Zephyr is most likely
going to be hit the hardest by it (most other services can have some
form of caching, but message delivery cannot).
When would it be convenient to get us all together to discuss the
problem? I do not think any solution proposed to date is adequate
(and this includes the client delivery mechanism that Jeff and I
quickly discussed over lunch one day). I think we need a brainstorming
session to throw out lots of ideas, and then one/two more meetings
after that when a concrete design is drafted (I'll take responsibility
for drafting such a design). (Perhaps the latter is something for the
TRB, but the overlap is rather great.)
I am unavailable:
July 6-9 (vacation)
July 30
I will also send out all the notes I have drafted up when I get back
from vacation, but for now, I simply want to schedule a time slot when
we can all sit down. Right now, I have brainstormed with a couple of
you one-on-one, but have been unable to draft a solid, workable
solution, which is why I want to call a larger brainstorming session
together.
-Richard
