🚀 go-pugleaf

An issue that I have seen kicked around briefly in these groups, and
now has arisen locally, is a question about the upper limits in Unix
for supporting large volumes of TCP connections and server processes.
There are many examples, but email is the one we are considering.  The
question is, given a Sparc 2 class machine and SCSI disk drives, how
many connections can it usefully hold open, and how many server
processes can it run simultaneously? Assume no users actually log in.

As a performance analysis problem, this clearly has many components,
including size of the process, duty cycle and traffic activity of the
client/server connection, amount of file activity per sever, and many
other things.  One way to answer the question is simply to try it with
a synthetic load and measure the response curves. Another way is to try
and understand the hard, intrinsic kernel limits, such as the upper
limits on file descriptors, the non-optimal searching algorithms for
TCP control blocks, routing tables and other heavily accessed network
data structures (eg. linear linked lists instead of hashing), process
table size limits, buffer limits, etc, etc, etc.

Has anyone tried to drive this class of machine hard in a
"connection-oriented" client/server relationship? Is there any relavent
performance analysis literature in this area? Or does anyone have any
fundamental insight into this problem based on their knowledge of
Unix?  Any comments on this topic would be welcome. If there are
vendors that have specifically addressed these issues in their kernel
implementation, I would like to hear about it.

 

-- 
Eric Carroll            University of Toronto Computing & Communications
                        Network & Operations Services, Network Development