Mailing-List: contact users-help@tomcat.apache.org; run by ezmlm
Precedence: bulk
Reply-To: "Tomcat Users List" <users@tomcat.apache.org>
Received-SPF: pass (athena.apache.org: domain of jjakub83@gmail.com designates
 209.85.212.46 as permitted sender)
MIME-Version: 1.0
In-Reply-To: <51855EBC.9070102@verizon.net>
References: 
 <CAJgUXc32DRf=JPawZeR211KfY9MaU+itjwiuW4oiR+nJcNoCfg@mail.gmail.com>
	<5184FA11.7040004@ice-sa.com>
	<CAJgUXc13OEz88rRy7bp-+U3YGLQ0jcO9+UGVGq_H16X2bedByg@mail.gmail.com>
	<51851A7D.4000601@ice-sa.com>
	<E2A1C832-E890-4DC6-B930-81C02335A2B6@gmail.com>
	<51854456.8050906@apache.org>
	<51855EBC.9070102@verizon.net>
Date: Mon, 6 May 2013 10:37:51 +0200
Message-ID: 
 <CAJqdOe8oBGPSBU47r0XsMMW==wLxNu=HkrsGZhwLpo3wJrAVjQ@mail.gmail.com>
Subject: Re: Designing for Load on TomCat
From: Jakub 1983 <jjakub83@gmail.com>
To: Tomcat Users List <users@tomcat.apache.org>
Content-Type: multipart/alternative; boundary=001a11c1e238c2c5ef04dc089d1c

--001a11c1e238c2c5ef04dc089d1c
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: quoted-printable

Yogesh

"with a page response time of 5 seconds"

is it current response time or acceptable await time?
if it is current response time, when was it measured, during max load, or
during min load ? how many concurrent requests where sent ?


why so long ? processor is busy, or communication with external systems ,
database, webservice, filesystem ?
I am not a specialist, but here are my doubts
  if you do processor intensive task setting much more threads than cpu
cores/hyperthreads may not help, it may even slowdown, question arises how
much more threads than cores is sensible value: cores +1, 2*cores ?
  if time is taken mostly by communication - sockets, filesystem, database,
webservices, my doubt is:
      assumption: tomcat uses nio - non blocking IO, so above operation
shouldn't block, so my question is why and when do we need more threads ?
isn't a thread doing other job when awaiting for NIO ?

I would be grateful if somebody more experienced could explain this, and
expand this topic.

regards
Jakub


On Sat, May 4, 2013 at 9:17 PM, David Kerber <dckerber@verizon.net> wrote:

> On 5/4/2013 1:24 PM, Mark Thomas wrote:
>
>> On 04/05/2013 16:01, Yogesh wrote:
>>
>>> Well my question is Is it a common design practice from your experience=
s
>>> to configure one node (maxthreads) for the scenario where all other nod=
es
>>> amongst which the load was distribued fail ?
>>>
>>
>> You design for whatever level of resilience you need to meet the
>> availability requirements.
>>
>> Mark
>>
>
> Which IME means allow for either one or two of the cluster nodes to fail,
> depending on how many you have to start with.  Never all but one, unless
> you only have two to begin with.
>
>
>
>
>>
>>> On the cluster part, wrt tomcats talking to each other do you mean the
>>> session replication feature or something else ?
>>>
>>> Sent from my iPhone
>>>
>>> On May 4, 2013, at 9:26 AM, Andr=E9 Warnier <aw@ice-sa.com> wrote:
>>>
>>>  yogesh hingmire wrote:
>>>>
>>>>
>>>>  On Sat, May 4, 2013 at 7:07 AM, Andr=E9 Warnier <aw@ice-sa.com> wrote=
:
>>>>>
>>>>>> yogesh hingmire wrote:
>>>>>>
>>>>>>  While planning / designing to build a web app that must scale to 20=
00
>>>>>>> concurrent users, distributed across 5 Tomcat nodes in a cluster,
>>>>>>> Apache
>>>>>>> at
>>>>>>> the front of course and the ability to serve 20 concurrent requests
>>>>>>> per
>>>>>>> seconds during business hours, with a page response time of 5
>>>>>>> seconds, how
>>>>>>> would we go about the ask ? What Apache / Tomcat / System (CPU/JVM)
>>>>>>> parameters should be considered for this design ?
>>>>>>>
>>>>>> I will provide the ABC, and leave the details for someone else.
>>>>>> You have 20 requests arriving per second, and it takes 5 seconds to
>>>>>> process one request and return the response.
>>>>>> So, over time, it will look like this
>>>>>>
>>>>>> Time   new requests   requests in-process  requests terminated
>>>>>>
>>>>>> 0        20              20                      0
>>>>>> +1s      20              40                      0
>>>>>> +2s      20              60                      0
>>>>>> +3s      20              80                      0
>>>>>> +4s      20             100                      0
>>>>>> +5s      20             100                     20
>>>>>> +6s      20             100                     40
>>>>>> +7s      20             100                     60
>>>>>> etc...
>>>>>>
>>>>>> So, in principle, and assuming nothing else is going on, you need 10=
0
>>>>>> concurrent threads in Tomcat to process these requests.
>>>>>> (I would take a healthy margin of security and double that).
>>>>>> Whether for that you need a cluster of Tomcats is another discussion=
.
>>>>>> And how much memory you need to allocate to your Tomcat(s) JVM(s) is=
 a
>>>>>> function of what your webapp needs, to process one request.
>>>>>>
>>>>>> The numer of concurrent users should be relatively irrelevant, if al=
l
>>>>>> you
>>>>>> mean by that is that some of these requests come from the same user,
>>>>>> but
>>>>>> they are otherwise independent of one another.
>>>>>>
>>>>>> Note that I have a suspicion that what you describe as "requests"
>>>>>> above
>>>>>> probably only count the requests to your webapp code, and do not
>>>>>> count the
>>>>>> additional requests for stylesheets, images, etc.. which may be
>>>>>> embedded in
>>>>>> any page that the user's browser eventually displays.
>>>>>> So unless you plan on serving those directly from the Apache httpd
>>>>>> front-end, you should take them into account too.
>>>>>>
>>>>> Thanks Andre and sorry for not mentioning about the other content tha=
t
>>>>> are
>>>>> actually requested by http get's from the jsp served.,
>>>>> There is quite a lot of ajax calls and static content and that can be
>>>>> served out of httpd, but as of now it is not. I know not the best way=
,
>>>>>
>>>>
>>>> but you can read the on-line documentation, I presume ?
>>>>
>>>> so i
>>>>
>>>>> assume i have to increment my thread count correspondingly..
>>>>>
>>>>
>>>> Well yes, because then you do not have 20 requests per second, you hav=
e
>>>> more.
>>>> Only you would know how many more, and how long they take to serve, bu=
t
>>>> the calculation is similar.
>>>>
>>>>
>>>>> While planning to threads on a single node, do i have to take into
>>>>> account
>>>>> the failure scenario where say all other 4 nodes fail and just this o=
ne
>>>>> node has to serve out the entire web app load. For that, do i have to
>>>>> provision the thread count as many as 4 times what i arrive for a
>>>>> single
>>>>> node ?
>>>>>
>>>>> Your thoughts?
>>>>>
>>>>
>>>> I think that you can figure that one out by yourself, no ?
>>>>
>>>> One more thing, to avoid you looking in the wrong direction : having
>>>> one Apache httpd front-end distributing calls to several back-end Tomc=
ats,
>>>> does not make it so that the Tomcat servers constitute a "cluster".  A
>>>> "cluster" is a name more usually used when the Tomcats are talking to
>>>> eachother.  In this case, they would not be.  It would just be the
>>>> connector, on the Apache httpd side, which distributes the load betwee=
n the
>>>> back-end Tomcats, and detects when one or more is not working anymore.
>>>>
>>>
>
> ------------------------------**------------------------------**---------
> To unsubscribe, e-mail: users-unsubscribe@tomcat.**apache.org<users-unsub=
scribe@tomcat.apache.org>
> For additional commands, e-mail: users-help@tomcat.apache.org
>
>

--001a11c1e238c2c5ef04dc089d1c--