Windows NT performance in switched networks
Ruth Reed
Hi. I'm a network administrator currently studying for my MCSE, and I've
run across a question and answer that I'm unsure about. I've been using
some of the "approved" testiing materials from various test prep companies
to do some exam simulations, and I ran across a question that I'm not sure
if I agree with. I have encountered inaccuracies and verifiably wrong
answers, which is somewhat troubling since one can't effectively study to
take a test where the answers designed by the testers are wrong to begin
with (thus, making all your study materials reflecting the correct answer
"wrong" for the test). Since the test preparers don't give out all their
wrong answers in advance, it's hard to prepare properly. :-) Hopefully,
these wrong answers are few enough to keep you from flunking the exam
because of your "wrong" answers. :-) However, in this case, I'm not
certain if their answer is wrong or my understanding is wrong, thus I'm
appealing to the NT and networking guru's among us to give me a good
step-by-step explanation of what happens in a given scenario. I would
appreciate your input very much because I am concerned about doing a good
job, regardless of whether or not I'm MS certified, and the answer to this
question directly affects my ability to do that because this is the basics
of networking infrastructure.
The scenario was given that an office had a single NT server with ten
employees connecting to the server via a 10 MBps hub. The server has 1 10
MBps nic. Every day at 4:00, the work/updates slowed down to an
unacceptable point as users did all their updates to the database at once,
and the users were unhappy. The assumption was given that the server had
sufficient processing power to handle the load (making the problem
exclusively network related). The requirement was to resolve the
bottleneck. The proposed solution was to install a 10 MBps switch with a
100 MBps port, install a 100 MBps NIC in the server, and connect the NT
server to the switch via the 100 MBps port. The correct answer to the
question was that the proposed solution did not resolve the problem.
This is something that I have a lot of questions about. Their question
didn't really give sufficient data to provide a good answer - performance
data for the network during the crunch period would have been helpful - the
extent of the bottleneck is a critical factor in determining the solution
since effective solutions will vary depending on the degree and severity of
the problem. How can one determine if a solution corrects a problem if the
problem is not completely defined so one can evaluate? Obviously, putting
a couple of NICs in the server would help (vs. just one), but depending on
the severity of the problem, it may not resolve it, especially in a hub
environment where adding an additional NIC could actually worsen the
problem by further dividing bandwidth.
The implied assumption here seems to be that if the clients are connecting
at 10 MBps to the switch, the switch is bridging to the server at about 10
MBps, making the bandwidth to the server a continued bottleneck. It also
seems to imply that 10 MBps users/servers connecting to a 10 MBps hub are
actually getting 10 MBps, making the addition of a 10 MBps switch the
equivalent of the existing infrastructure. This would be true in a single
user environment - the switch could only bridge traffic to the server at
the rate the user could send it, and the hub would allocate all available
bandwidth to the two connected users. However, I had understood that
switch architecture bridges across the backplane (thus the advantage of
very high bandwidth and low latency), which theoretically would allow all
10 MBps clients to bridge to the 100 MBps connection at once due to
packetization (minus some overhead, since you don't get a full 100 MBps -
perhaps about 60 - 80%, cut in half by return traffic during peak loads -
still a respectable doubling of bandwidth to the server - say 20 MBps for
argument's sake - certainly noticeable by the users, and the equivalent of
adding a second NIC to the 10 MBps network and eliminating the hub
bottleneck of concentrated 10 MBps max. User bandwidth is also affected
thus, so the ratio of improvement should be the equivalent of the actual
stated bandwidths). If the switch is just that - only switching, and
single point/port bridging, it should still significantly increase
available bandwidth because the clients wouldn't be divvying up 10 MBps -
each would get the full 10 MBps when their turn comes around, and should
still observe a noticeable improvement in throughput. The server would
also have a lot more bandwidth, and thus the clients wouldn't have to wait
to get their "time" with the server since the server bandwidth could take
the 10 MBps transmissions from the clients as quickly as they can send them
rather than giving each of fraction of the server's 10 MBps or forcing
clients to wait. I'd also argue that networking is packetized, so these
communications shouldn't be "connection" bottlenecked - packetization
should afford some additional throughput simply because packets allow more
efficient bridging of traffic and would reduce timing delays due to slower
bandwidth on the concurrent 10 MBps client connections.
I had understood that networking was multitasking (a secondary
communication may commence at the NIC prior to the completion of an
original communication to a client that may be delayed because of network
congestion or latency) - thus, in a multiuser environment, the bridging
back to 10 MBps shouldn't be a major bottleneck as long as multiple users
are communicating with the server at the same time, which the scenario
stated was the cause of the problem and thus the solution can assume a
concurrent multiple user environment. Switching should allow the server to
concurrently communicate with more than one user at a time - parallel
communication vs. serial - which would allow continued high bandwidth on
the return path despite the latency involved in bridging back to a slower
10 MBps. Communications should still increase in speed due to bridging and
the higher bandwidth over the backplane and to the server, and latency
shouldn't become so great as to require frequent packet retransmission.
I'm not sure about switch buffering (i.e., what happens if communications
exceeds the bandwidth available), and how much data the switch can hold
prior to bridging to a slower connection when switch bottlenecks occur.
I'm not certain exactly how much latency is involved in bridging at the
switch (and it varies, depending on manufacturer), but it shouldn't
completely offset the advantage of the additional bandwidth. I'm also
confused as to why the assumption was made that communications via a hub
(which concentrates) is the equivalent of communication via a switch (which
uses the backplane bandwidth to bridge). For example, adding a second nic
to a server in a hub environment wouldn't resolve the problem, either.
Adding a second NIC in a switch environment would effectively add
additional bandwidth, incrementally improving the problem, which is what
most real world network solutions are (since the corporate brass tends to
get huffy if you suggest yanking out all the existing obsolete and
inadequate network infrastructure and replacing it with what really works).
The proposed solution is also an incremental improvement. My (possibly
inaccurate) assumption is that in a hub environment, you're only getting a
fraction of your 10 MBps bandwidth anyway, since I had understood the max
capacity of a 10 MBps hub was 10 MBps, and each connected user/server gets
their slice of that (vs. the alternative of 10 MBps each that one might
think a hub would give them based on the speed of their connection to the
hub). Without knowing the extent of the proposed problem, how can I
determine whether or not the incremental fix proposed will solve it? The
"solution" absolutely depends on the degree of the problem - adding in a
100 MBps switch(es) and migrating all users to 100 MBps wouldn't solve a
severe problem despite the exponential boost in bandwidth - you'd need to
add extra nics to the server, subnet, reroute and such as well. Does NT
have networking limitations in a switched environment that I haven't
considered? I find this question very troubling.
If anyone out there can go over the technical process of a switched network
via NT with me step-by-step and describe how and why the bottlenecks along
the way occur, I would appreciate it. I'm sure there's some inaccuracies
in my understanding of switched networks, and I definately want to
understand this better. It's one of those nagging questions, you know?
I'll try to get an actual copy of the question and the name of the test
preparer for you, if you'd like.
For all those MCSE's out there, has anyone seen this question on an actual
MS exam? For all you Microsoft folks, can you doublecheck your exam
scenarios and the accuracy of the proposed solution and correct answer for
this type question?
Please respond by e-mail, since I don't know if everyone in these
newsgroups are interested in switched NT networks.
Thanks much!
Ruth Reed
rr...@ddc.net
Gentry Howard
Ruth,
Is it possible that they we're leading you in the wrong direction with the
proposed soultion? Given the problem I would look at tuning the database
(assuming a SQL server on the NT server...) if the problem was too many
database updates at once.
I can't help with the switched network question as I too have questions as
to exactly how this works (with btw you put rather well in your original
post...)
In reality land we had similar problems with a non C/S database, upgraded
all the clients to 100mb, as well as the server(not an NT server...) and the
problem was relieved, or at least returned to acceptable tolerances.(Which
may be the point of the problem in saying that the proposed soultion did not
solve the problem...)
Needless to say we're migrating to c/s apps....
Ruth Reed wrote in message <01bd3fa6$b699dda0$a537...@rreed.ddc.net>...
Craig Paul
On the NT server in question I'd fire up PERFMON during the time that
response was fine and watch both the CPU utilization, disk utilization, and
most importantly, the page file/swap file statistics and the amount of
paging that's happening. Then compare those stats to the stats you get at
during the jam and see what the bottleneck is.
PERFMON is found in Start/Programs/Administrative Tools/Performance Monitor
Ruth Reed wrote in message <01bd3fa6$b699dda0$a537...@rreed.ddc.net>...
Mike Crabtree
Unless you can quote the question verbatim with all punctuation and
annotations intact I couldn't really comment on it's validity..
These questions are specifically designed to catch you out - and very subtle
pieces of information often result in a totally different answer...
Having said that - yes - in my experience, a few are wrong.....
Mike Crabtree MVP - email: mike @ imjc . com
Ruth Reed
See below for entire question text.
After re-reading the question, I realized that my initial recollection of
the question was not quite correct (it had been several days since I'd
played with the test prep software). However, I still feel the question is
not properly formulated to allow an individual to derive an informed
response. One does not have sufficient facts regarding the problem
situation to determine whether or not the given solution corrects the
problem. Server performance issues can have a multitude of underlying
causes, and nothing in this scenario points to any individual problem cause
or rules out any potential solution. No specs for the server are given, no
data regarding the transaction software (if any) are provided, and no
network related utilization data is included. For all we know, the server
could be a 286 with 2 megs of RAM trying to run NT 4.0 (an exageration
since NT 4.0 couldn't run on that thin a machine, but you know what I
mean). Or the clients could be transfering files back and forth to the
server - I've personally witnessed via protocol analysis a single client
hike 10BaseT network utilization to 94% during file transfers, so the
number of clients connected to the server tells us absolutely nothing about
network utilization. We don't know what protocols they're running
(Netbios/Netbeui could be a problem if it's generating a lot of broadcasts
during the critical period) or whether any of the NICs are chatty. We don't
even know what NOS they're running.
My vote for this is -
The correct answer is "E. Insufficient data to determine appropriate
response."
This has made me realize how difficult it must be for Microsoft to
formulate scenario based questions that provide sufficient data and
complexity to approximate a real-world network problem and still allow the
testee to determine an experience based answer. Yikes!
I'd be interested to hear your opinion on this question. I would also
still like to hear an explanation of switched network bottlenecks and data
flow so I understand it better. That way, if I get a "real" switched
network question - either in the real world or on an exam - I'll know what
to do.
To be fair to this particular test prep company, I've also found obvious
errors on the transcender demo's, so I know they all make mistakes. But -
here's a quote from TopGrade test preparation software - the duplicate a/b
answer is on the software:
Situation:
You are hired as a network consultant at Springfield Toys to address their
Server response problem. Springfield has one file server and several nodes
on a single 10BaseT Ethernet LAN segment. During the evening hours between
4 pm - 6 pm, when every one tries to update their daily transactions on
the Server, the Server response degrades to an unacceptable level.
Required Result:
Improve the server response time
Desired Optional Results:
Minimize Costs.
Proposed solution:
Install a Fast Ethernet Switch with one 100 Mbps port and 4 10BaseT ports.
Connect the File Server on the 1000 Mbps port and distribute the other
nodes equally between the 4 10BaseT ports.
a) The proposed solution produces the required result and the optional
result.
b) The proposed solution produces the required result and the optional
result.
c) The proposed solution produces the required result but not the optional
result.
d) The proposed solution does not produce the required result.
Brian Steele
Ruth Reed wrote:
> Situation:
> You are hired as a network consultant at Springfield Toys to address their
> Server response problem. Springfield has one file server and several nodes
> on a single 10BaseT Ethernet LAN segment. During the evening hours between
> 4 pm - 6 pm, when every one tries to update their daily transactions on
> the Server, the Server response degrades to an unacceptable level.
>
> Required Result:
> Improve the server response time
>
> Desired Optional Results:
> Minimize Costs.
>
> Proposed solution:
> Install a Fast Ethernet Switch with one 100 Mbps port and 4 10BaseT ports.
> Connect the File Server on the 1000 Mbps port and distribute the other
> nodes equally between the 4 10BaseT ports.
>
> a) The proposed solution produces the required result and the optional
> result.
> b) The proposed solution produces the required result and the optional
> result.
> c) The proposed solution produces the required result but not the optional
> result.
> d) The proposed solution does not produce the required result.
I think that the key here is that we're talking about a
transaction-processing environment, and not a file-serving environment.
For example, I manage a REAL transaction-processing environment here,
with a VAX 7610 and 100+ users. Network traffic in this type of
environment can be really low, depending on the applications involved.
In my case, network traffic to/from the VAX averages 25-26 KBytes/sec
during the day, peaking at 75 KBytes/sec at times. This is well within
the limitations of a simple shared-ethernet scheme, and performance
problems will most likely be caused by server limitations and not
network congestion.
I would therefore select (d) as my answer.
Brian Steele
IS Manager
GRENTEL
Ruth Reed
Hi - thanks for the info. Yikes! Then can you please explain it to me?
Specifically, tell me about the transaction environment you're in and how
it functions - where the processing happens, what's actually passed from
the client to the server, etc. I believe you've hit the nail on the head
for me - I've never worked with transaction servers (except NT or Novell
type servers, which could suffer from both server and bandwidth
bottlenecks), and most of what I've been doing has been client/server or
internet related. What kind of clients are accessing the server for
transaction updates? How does the transaction process limit network
bandwidth utilization since even simple database updates from a client can
generate a significant amount of network traffic? Is transaction computing
centralized vs. client server? Can the assumption be made that all
transaction environments involve low network utilization? For example,
could I say without a doubt that the problem given in the sample question
is definately server related, or is there still some doubt based on
circumstances and technology utilized? Assumptions during troubleshooting
can be "dangerous" and expensive. Since I haven't worked in a transaction
environment, I'd love it if you would give me a bit of info on data flows
and such. However, I will stick with my "E." response since I've
personally seen environments where either/or could be a cause - selection
of the server as underlying cause is itself an assumption since for all we
know it could be a supercomputer (since the info isn't given.
For the record, I've worked with computers for 15 years and have plenty of
experience, just not with transaction servers of this type, as you say.
(oops!) Which is exactly why I needed clarification of the question and
some assistance from the nice fellow technical folks in this forum. I have
hopes that someday I'll know everything there is to know about every kind
of technology, but I suspect it won't happen any day soon. And if I ever
get real close, they'll change technology enough to make me start over!
:-) So no matter how much I learn or how long I work and how much I know,
I will forever be a bumbling idiot trying to figure out how things work.
Thanks again for your reply, and I look forward to a bit more info so I
understand the topic properly. Or, if you're too busy for the explanation,
if you know of some links to good web sites or white papers describing the
topic that you could e-mail to me ...
Brian Steele <omi...@caribsurf.com> wrote in article
<34F41F...@caribsurf.com>...
Greg Thomas
I think y'all are reading too much into this question. I believe
they are looking for a way to increase bandwidth to a basic
file server without replacing all the network cards.
Ruth Reed <rr...@reedcomputer.com> wrote in article
<01bd421c$a62fa620$9837...@rreed.ddc.net>...