hc-httpclient-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Roland Weber <http-as...@dubioso.net>
Subject Re: Does HttpClient create connections on fly
Date Sun, 18 Dec 2005 12:06:12 GMT
Hi Bharathi,

> We will use only one httpConnection and MultiThreadedConnection manager > for the
total application.

I assume you mean one HttpClient, not one HttpConnection.

> But inorder that all the threads in case 160 are launched at the same time,
> get back the response in 3 seconds, do i need to add any thing extra in code
> by making the connection persistent.

We don't know your code, so we really can't tell what you need to add or not.
HttpClient supports connection re-use out of the box if used correctly, but
your other comments suggest that your application is not using it correctly,
or that your environment does not allow connection re-use. For example, if
the server sends back "Connection: close" headers with each response, it is
not possible to re-use the connections.
The most common error that prevents connection re-use is that people create
a new HttpClient object and/or HttpConnectionManager for each request they
send. That was the first thing we asked you. Now that we know you are not
making that mistake, we need more information from you to learn why you are
not seeing connections being re-used. A wire log of two requests being sent
to the same server consecutively (*not* in parallel) would be a good start.
Feel free to remove the request contents from the log, what we need is the
headers and traces from HttpClient code.

> The production boxes have 8 cpus. I need the values to be set for the
> maximum connections per host.

There is a step called "performance tuning" that should be part of any
application deployment into a production environment. You develop a load
test and run it against some initial values for your tuning parameters.
Then you vary some of the tuning parameters and run the load test again
to see if there is an improvement. Repeat this until the performance is
satisfactory. If you can't achieve satisfactory performance though the
hardware is powerful enough, review the application code to identify and
remove bottlenecks.
If you are hitting only one host, set the maximum number of connections
to the same value as the maximum number of connections per host. You've
got 8 CPUs, so there really is no point in starting with less than 8
connections, as a rough guess. If you find that performance can no longer
be improved by increasing the connections per host, but the CPU load is
still low, then consider vertical scaling of your application.

We know how HttpClient works, but we don't know your environment nor your
application. We don't know how many requests you have to send in a given
timeslice, we don't know how much data is sent with each request, we
don't know how much data is returned by the server with each response,
we don't know how much time the server needs after receiving a request
before it starts to send the response. Neither do we know about possible
design flaws in your application, or peculiarities of your environment.
If you have only one HttpClient, are you supplying individual HttpState
objects for each request or do you share the default state across all
threads? Do the servers use HTTP/1.0 or HTTP/1.1? Are there proxies or
firewalls inbetween? There are a lot of things that could have an impact
on HTTP performance in general, and on HttpClient in particular. We are
in no position to give you any reliable values for tuning HttpClient.

> I alos need a clarifictaion do i need i need to add any thing extra for
> pooling and closing the connection. In the test i did i find the
> connections are not being reused at all.

This is a specific problem which we can help you to solve. Send us a
wire log (http://jakarta.apache.org/commons/httpclient/logging.html)
and somebody is going to have a look at it. Some code fragments that
show how you are configuring and using HttpClient might also be helpful.

> But how do i know what are the values to be set. The no of threads
> launching is not controlled by us. It might be 150 or more. But how
> do i decide the max no connections for host and total connections.
> That is, which one is faster, running the same with lesser no of
> connections in pool or having 150 in the connection   pool.

Develop a load test for simulating varying numbers of threads. Consider
limiting the number of threads by some external means. Since you have
no upper bound on the number of threads, you can never guarantee that
there is a connection for every thread available. Besides, connection
re-use is likely to decrease with the number of connections, right?
You should try to simulate a high-load situation as closely as possible,
with real-life behavior for the amount of data being transferred, the
response times from the server, and the frequency in which the threads
generate requests. Then play around with the number of connections.

> I also want the clarifictaion if the connections are being reused.

They are not, you told us yourself. We don't know why.
You should fix that before running performance tests.

> We want the response back in 3 secs,but if the aplication connectivity
> itself is taking 1 minute then it is a big problem.

You can set timeouts to report errors instead of locking up threads.
You can implement your own connection manager to detect specific
load situations and report them, at least in a development or system
integration environment. You can implement your own connection manager
to create new connections if the delay until an old one is released is
getting too long. You can clog your application with high load, then
send signal 3 to the JVM to generate a core dump, which will tell you
where exactly all the threads are really waiting. (I would try that
before implementing a custom connection manager.)

Personally, I think you have some very basic flaw in your application
that prevents the connections from being re-used. Your first step
should be to locate and remove that flaw. But we can't help you if
all you're telling us about the problem is that you have it, while
asking for performance tuning values to scale your application to
an 8 CPU box though it doesn't seem to scale on a single CPU.


To unsubscribe, e-mail: httpclient-user-unsubscribe@jakarta.apache.org
For additional commands, e-mail: httpclient-user-help@jakarta.apache.org

View raw message