Maximum Concurrent TCP Limitations in Real World

gsgi, thank you.  The numbers you mentioned on exchange server are interesting.  Any idea what kind of hardware they are using for that load of 10K per machine?

Anyone else have any data?

Well that was five years ago, and I believe they were 4 processor pentium 200 machines running advanced server with about 2gb's each.

Here's a rough idea of Experts Exchange setup: "There's a database server with a redundant backup. There are eight on-line webservers, with backups. There are two email servers dedicated just to sending notifications and nothing else."  So that is 10 servers for 60,000 active users of the 3 million users.   Do you see a bottleneck?  1 Database server (no bottleneck) - 2 email servers sending e-mails (different than my example above which was users not dedicated to sending only) - 8 web servers!  So you see, it is the *task* that matters as much as the number of connections.  Even then though, if there really are 60,000 active users, then there are 6,000 users per server if there are really 10 servers.  That's lower than my example, but we see a web server bottleneck.  Same with hotmail - its a farm of servers.  By the way, you can run a farm of windows servers (like 10) run on one physical box - its called a vmware solution.  I think they put it on a IBM RS 6000 machine.

-gsgi

Do you think the numbers would scale then based on the Exchange server using 5 year old equipment.  Say a top of the line server nowadays for example.  Do you think it would then do 10 times the user limit?  I also wonder if there are OS level limits in Windows that you would reach first?

Not for windows.  For windows maybe x 2 but for linux, apache, mysql, maybe better -

This guy is at 1 server per 50,000 users - I would be less optimistic.  Although at 4% usage, see the post below the link, that's only 2,500 concurrent users per server.  Even this optimist has more content (web) servers than database servers.

http://www.experts-exchange.com/Storage/Q_21400842.html?query=%22how+many+servers+for%22+users&clearTAFilter=true

-------------------
 Accepted Answer from infotrader
Date: 03/18/2004 01:09PM PST
      Accepted Answer       

Also, it's not likely that you'll have all of your clients accessing the server at the same time.  According to past experience, you should normally expect 1 to 4% of users... Typically, if your user base is 100 people, maybe only 2 would be watching at any given time.

- Info
--------------------

Thank you for your help.  I suppose I will only really know after testing my application since its unique to what others are doing.  Web servers are designed to kick you off as soon as possible, whereas my server is meant to hold the connection until the user leaves in hours or days.  Webserver statistics I find to be not very useful because of this.

The only other thing I am looking for right now is to know if there are OS limitations on 2K/2K3 (Server Editions) and if there is anything I can do about them if they exist.  No point in desiging something that won't work no matter what if you see what I mean?

Why would you use Windows for this?  Why not put it on a Linux Box?

Actually our server code is written to be portable between Linux and Windows.  I want to know for both platforms what I can expect to achieve in a realistic scenario, but it will be getting a lot more Windows use initially.

how many users do you have ready to test it - you could throw it up at shared hosting site and blast away. - gsgi

The project is in secret development at the moment and not ready for the public.  We have some pretty high end servers on multiple 100mb pipes that we are testing internally on.

We are writting a load-test application that we can run from many other servers to try to simulate a lot of users, but once again, I was looking for anyone's existing real-world data and possible OS limitations.  I don't want to spend months trying to optimize somethign that is not in my code/design/control.

Here some stuff I found:  again, I am seeing between 10,000 and 100,000 users per server depending on what it is doing and how efficient the code goes.
And again, web serving is at the bottom of that.

http://netwinsite.com/dmail/perform.htm

http://office.microsoft.com/en-us/assistance/HA011647621033.aspx

http://www.google.com/search?hs=Vc9&hl=en&lr=&safe=off&client=firefox-a&rls=org.mozilla%3Aen-US%3Aofficial_s&q=hardware+requirements+%2210%2C000+users%22&btnG=Search

 System Requirements - MailMarshal for SMTP

Hardware and Software:
Hardware requirements naturally vary depending on the number of email users and the amount of email traffic. The following minimum specifications are suggested as a guideline:

1,000 Users:

    * Pentium 1.8Ghz
    * 10Gb Hard Disk Drive*
    * 512Mb Ram

10,000 Users:

    * Dual Pentium 1.8Ghz
    * 60Gb Hard Disk Drive*
    * 1024Mb Ra

gsgi,

My only problem with this information is that it's very hard to show how many "concurrent" users we are talking about in any of those scenarios.  Your first example of Exchange Server was probably the best because I believe it does keep a constant connection open to the server.  Most other mail platforms only "poll" the server occassionally.  If I read this right D-Mail for example limits you to 200 concurrent by default.

Regardless, Thank you for your help and I will for sure give you some of the points.

OK, but everybody knows that Exchange is not the fastest.  Groupwise, by Novell, is probably a more realistic measure in the e-mail world, and they are moving to Linux for their platform and most of their customers (left) are really big - 10,000s of users.  - good luck - gsgi

Anyone else? Still looking for futher real-world examples and/or information about OS Limitations.

Thanks.