tomcat-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Andrew Skiba" <ski...@gmail.com>
Subject Re: backlog measurement
Date Fri, 11 Jan 2008 08:24:30 GMT
Hello Filip,

Thank you again for your reply and for the links.

I think that makes sense.

Yes, I will end up with unmeasurable backlog after acceptQueue is full. But
when the queue is empty, which is the normal operating mode, the backlog
will be filled only for few milliseconds of context switch. AcceptRunnable
is only accepting sockets and putting them into the queue, so it's hard for
kernel backlog not to be empty.

Anyway, the purpose of this patch is not to measure backlog size per se, but
to identify a server incapable to handle its load for long periods of time -
tens of minutes, or hours.

You are right, different OSes have tools to monitor this, but it's always
good to have a pure java solution at hand.

In underloaded server this patch will add an overhead of 1 thread per port,
and rarely few extra I/O buffers for short periods of higher load. I
suppose, GC will take care of that, when the load is low again.

If the server is slashdotted, you will know this sooner or later. The worst
situation is when the load is just slightly above the possible to handle. If
I have 10 worker threads and 20 simultaneously stressing clients, the
response time is double of what can be measured on servlets. This patch
allows to figure that out at price of 1 additional thread and 10 extra I/O
buffers. Probably there will be people who are willing to pay this price, so
they will use this connection factory.

Regards,

Andrew.

On 1/10/08, Filip Hanik - Dev Lists <devlists@hanik.com> wrote:
>
>
>
> hi Andrew, the solution is a bit overkill and you may be
> misunderstanding the backlog concept.
>
> the concept behind the backlog, is when the app is too busy accepting
> connection, the kernel and its TCP stack will handle it for you.
> and by doing this, you can balance the pressure of new connections at
> the kernel level.
>
> for example, if all threads are busy in java handling requests, the
> kernel will handle SYN and TCP handshakes until its backlog is full.
> when the backlog is full, it will simply drop future SYN requests. it
> will not send a RST, ie causing "Connection refused" on the client,
> instead the client will assume the package was lost and retransmit the
> SYN. hopefully, the backlog queue will have cleared up by then.
>
> doing this in java, and by accepting connections by the process running,
> you're doing redundant work, hence it would be overkill.
>
> in terms of backlog reporting, which is the idea behind this, one would
> have to look at the specific OS, and see what it has to offer for that
> kind of info.
>
> a great read to understand the backlog functionality, since its so
> loosely defined is
>
> http://www.amazon.com/Unix-Network-Programming-Addison-Wesley-Professional/dp/0131411551/ref=pd_bbs_sr_1?ie=UTF8&s=books&qid=1199991167&sr=8-1
>
> it explains it fairly well.
>
> in your example, maxThreads=10 acceptCount=30, with your
> implementations, your not measuring the actual backlog, you're only
> measuring the number of actually accepted connections, you still end up
> with unmeasurable backlog. what you instead end up with is 15 extra
> connections, and the TX/RX buffers and the java objects associated with
> it.
>
> does that make sense?
>
> Filip
>
> Andrew Skiba wrote:
> > Hello Filip, thanks for your reply.
> >
> > You are absolutely right.
> >
> > I did this trick when I understood that it's not trivial to measure
> backlog
> > in Java. I called it backlog because a normal system would have backlog
> of
> > this size in the same conditions. For example, for maxThreads=10 and
> > acceptCount=30 and 25 client threads the backlog size will be 15,
> although
> > we cannot measure it.
> >
> > If you use by socket factory in the same conditions, it will report
> > "backlog" size=15. Although now these 15 sockets are in totally
> different
> > state, from logical point of view, my blackbox behaves the same and now
> is
> > able to report this number.
> >
> > About memory consumptions. In normally functioning system I expect
> threads
> > to be blocked on accept() 99% of the time. Backlog is a spare for short
> term
> > spikes. So my implementation will accept a socket and this socket will
> be
> > used by an other thread very soon. Also it's possible to configure a
> half of
> > a usual acceptCount because I create an additional buffer of this size.
> >
> > I intend to use this socket factory to monitor and store the "backlog"
> size
> > once in a few minutes for all my servers. If I see that in last 3 hours
> this
> > size is always above zero, I know there is a problem. Or if users
> complain
> > that between 17:00  and 17:30 my application responded slower than
> usual, I
> > can check what were backlog readings at that time. Currently I can check
> > only timestamps when requests were processed, but requests could be
> delayed
> > in the backlog and I have no way to know.
> >
> > If my solution is an overkill for this problem, can you please advice me
> a
> > more appropriate way.
> >
> > Thank you
> >
> > Andrew.
> >
> > On 1/10/08, Filip Hanik - Dev Lists <devlists@hanik.com> wrote:
> >
> >> I'm a bit confused on how you can measure backlog in Java. Backlog is a
>
> >> TCP stack implementation setting.
> >> Also, between TCP implementations, there is no firm definition of what
> >> backlog actually means. does it mean SYN_RCVD or ESTABLISHED but not
> yet
> >> accepted?
> >>
> >> If I read the implementation correct, this has nothing to do with
> >> backlog, since you are accepting the connections into the process, ie,
> >> the handshake is done, and the accept call has returned. this
> connection
> >> is no longer subject to the TCP backlog (IIRC).
> >>
> >> This implementation, would prematurely accept connections, causing the
> >> system to take up memory for send and receive buffers for the Java
> >> process.
> >> Filip
> >>
> >> Andrew Skiba wrote:
> >>
> >>> Hello,
> >>>
> >>> I want to contribute a custom SocketFactory allowing to analyze the
> >>> utilization of acceptConnection attribute of a Connector. In a
> >>> properly configured production system, there should be rare situations
> >>> where connections wait for a worker thread to be handled. Our client
> >>> complained on high latency of web requests, but the measurement on
> >>> servlet did not show high latency. So we wanted to know the number of
> >>> connections which wait in socket backlog and were not accepted yet.
> >>>
> >>> I solved this problem by writing a custom SocketFactory, which accepts
> >>> connections immediately and puts it in my queue until a call to
> >>> accept() will take them. So the number of waiting connections can be
> >>> monitored via JMX.
> >>>
> >>> To activate this factory, the declaration of the corresponding
> >>> Connector in server.xml should be changed like in the following
> example.
> >>>
> >>>  <Connector port="8080" maxHttpHeaderSize="8192"
> >>>                maxThreads="10" minSpareThreads="5" maxSpareThreads="7"
> >>>                enableLookups="false" redirectPort="8443"
> >>>
> >> acceptCount="10"
> >>
> >>>                connectionTimeout="2000" disableUploadTimeout="true"
> >>>
> >>>
> >>> socketFactory="
> >>>
> >> org.apache.tomcat.util.net.BacklogMeasuringServerSocketFactory"/>
> >>
> >>> No changes in existing classes are required.
> >>>
> >>> Please review the code in the attachment.
> >>>
> >>>
> >>> Andrew Skiba.
> >>>
> ------------------------------------------------------------------------
> >>>
> >>> ---------------------------------------------------------------------
> >>> To unsubscribe, e-mail: dev-unsubscribe@tomcat.apache.org
> >>> For additional commands, e-mail: dev-help@tomcat.apache.org
> >>>
> ------------------------------------------------------------------------
> >>>
> >>> No virus found in this incoming message.
> >>> Checked by AVG Free Edition.
> >>> Version: 7.5.516 / Virus Database: 269.19.0/1216 - Release Date:
> >>>
> >> 1/9/2008 10:16 AM
> >>
> >>
> >>
> >>
> >> ---------------------------------------------------------------------
> >> To unsubscribe, e-mail: dev-unsubscribe@tomcat.apache.org
> >> For additional commands, e-mail: dev-help@tomcat.apache.org
> >>
> >>
> >>
> >
> >
> > ------------------------------------------------------------------------
>
> >
> > No virus found in this incoming message.
> > Checked by AVG Free Edition.
> > Version: 7.5.516 / Virus Database: 269.19.0/1216 - Release Date:
> 1/9/2008 10:16 AM
> >
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@tomcat.apache.org
> For additional commands, e-mail: dev-help@tomcat.apache.org
>
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message