How can I benchmark TCP and UDP using C?

views:

484

answers:

+3 Q:

How can I benchmark TCP and UDP using C?

I want to measure message latency and throughput for both TCP and UDP using C sockets. I am writing a client.c and server.c (Question 1: Do I have to?) that run on different servers to accomplish this. I have to take several measurements using different packet sizes and multiple trials and plot my results.

For message latency: Send a packet of varying size and measure the round trip time and divide by 2 to get latency. Question 2: I send a packet of size x, and then from server.c I send the packet back. So at the client I start the timer, send packet and then wait till I receive the packet then stop the timer. The timer/2 is my latency? What are the steps to go about measuring this?

For throughput: Question 3: How would I measure throughput of both? What are the steps to go about doing this?

Im a beginner to Unix socket programming in C so detail would be helpful, with emphasis on the bechnmarking aspect.

EDIT: I cannot use tools that already exist. I need to write my own.

Thanks!

+1 A:

Question 1: You probably can find existing test cases with some use of Google. But if there are special situations you need to deal with, then it probably makes sense to write your own so that you take that into account.

Question 2: Maybe there is an official definition of "latency" in terms of network operations, but I think what is really important is the round trip cost. So I would not divide by 2. The total round trip cost is what the user (client) experiences. That will be the latency they see.

Question 3: I think your plan of using differing packet sizes is good for measuring the throughput. There is another SO question related to this. I posted an answer to that question and provided some numbers from my own testing of UDP versus TCP. Those might be of interest as a "sanity" checkpoint.

Ah - I forgot one thing I was going to mention. If you write a really simple test case with UDP, it might be somewhat unrealistic. If you use UDP directly, you will need to add your own error handling, packet verification, etc. That will add cost. Ultimately, we find that UDP is faster in most situations for us because we have tailored it to our own use. But it certainly requires a LOT more code to get everything working correctly.

Mark Wilkins 2010-02-27 21:21:35

For question 2: How would I go about doing that? my Server.c would just grab packet and resend it where it came from? Is that how I test round trip time?

NewSocketUser 2010-02-27 21:27:27

@NewSocketUser: Yes - that is what I would do. With UDP, for example, the recvfrom() call (if I remember correctly) returns the data as well as the network address of the sender. You can then just send it back. But UDP is trickier since there is a maximum packet size that can be sent. So you basically need to do a recvfrom call for each packet sent (so you probably need to send data length in the first packet).

Mark Wilkins 2010-02-27 21:44:05

So round trip time is done on a per packet basis and throughput is done on a per message basis, would could take up to multiple packets?

NewSocketUser 2010-02-27 22:41:11

It probably depends on the application and how you define it. I think that most network conversations that involve large volumes of data are primarily sending the bulk of the data one direction (e.g., streaming video, writing a huge image to a database, etc.). So you would probably want to measure throughput with sending large volumes one direction ... just my opinion based on usage scenarios I am familiar with.

Mark Wilkins 2010-02-28 00:37:33

+2 A:

There already exists a program called ttcp which test TCP and UDP performance:

DESCRIPTION

Ttcp times the transmission and reception of data between two systems using the UDP or TCP protocols.

It does not calculate latency, but you can take a look at the source code of that program to examine how it does other calculations. And you can also use that as either client or server in case you do not want to write both yourself.

Update: Other similar TCP test programs

as well as related programs

hlovdal 2010-02-27 21:26:52

So you're saying I can use its source code as a starting point?

NewSocketUser 2010-02-27 21:39:09

I cant use an already built solution I have to write my own

NewSocketUser 2010-02-27 21:39:33

Even when you write your source from scratch, you can still use ftp://ftp.sgi.com/sgi/src/ttcp/ttcp.c as a reference to see how things are done.

hlovdal 2010-02-27 22:32:30

You may want to look at tools that already exist for this type of thing. Maybe D-ITG or iperf combined with tcpdump or wireshark? Unless the intent is to learn more about socket programming itself.

ezpz 2010-02-27 21:33:59

You might also be interested in this related question:"How to measure network performance (how to benchmark network protocol)".

none 2010-03-10 20:25:21

ansaurus

tags:

views:

answers:

How can I benchmark TCP and UDP using C?

related questions