Menu
▾
▴
Re: [GridSim-users] Question regarding simulation time
|
From: Chainan S. <ch...@sc...> - 2007-12-28 09:28:18
|
Good morning Agustin,
First, thank you very much for your prompt reply. I think I'm understand
about start time of the job now but I am still in doubt about the reason
makes the lower bandwidth user having its gridlet ready for execution
earlier than higher bandwidth. I am sorry that I think I was not able to
explain my questions clearly last time. Let me try to put those
questions again. I will put the program code and result here for ease of
discussion.
----QUOTE (TestCase1.java)----
// sends all the Gridlets
Gridlet gl = null;
boolean success;
for (i = 0; i < list_.size(); i++)
{
gl = (Gridlet) list_.get(i);
// For even number of Gridlets, send with an acknowledgement
if (i % 2 == 0)
{
success = super.gridletSubmit(gl,
resourceID[index],0.0,true);
System.out.println(name_ + ": Sending Gridlet #" +
gl.getGridletID() + " with status = " + success +
" to " + resourceName[index]);
}
// For odd number of Gridlets, send without an
acknowledgement
else
{
success = super.gridletSubmit(gl, resourceID[index],
0.0,false);
System.out.println(name_ + ": Sending Gridlet #" +
gl.getGridletID() + " with NO ACK so status = " +
success + " to " + resourceName[index]);
}
}
----End of QUOTE----
>From the code above, will each user submit all gridlet to
resourceID[index] at the same time? Let's say 4 gridlet per user in this
example, should they all send to execute at the same time? Or the
submission of job is hold until the termination of earlier gridlet?
My understanding is if it is even (ack is required), then, the
submission of next gridlet is hold until the first gridlet already
submitted (it probably haven't started its execution yet but the job is
already transferred along with its in_file.) Otherwise (no ack), the
next gridlet is submitted immediately (the gridlet may not yet finish
transferring the in_file while new gridlet is submitted) Am I correctly
understood?
If so, the users bandwidth should have an impact on the start time (I
mean the time that user starts to submit the gridlet, not the starts of
execution of gridlet) and duration of the gridlet's submission. Lower
bandwidth leads to longer time to make the gridlet into the "ready for
execution" state, and also delay in submission of subsequent gridlet (to
be more specific, gridletSubmit causes longer delay, if it requires ack)
Now, let's back to Gridlet's result. I copied only first Gridlet of both
user here. Please notice the line with **** is highlight the submission
time.
------QUOTE - User0 - Gridlet 1 result -------
Time (sec) Description Gridlet #0
------------------------------------------
0.00 Creates Gridlet ID #0
0.00 Assigns the Gridlet to User_0 (ID #17)
11.00 Allocates this Gridlet to GridResource_0 (ID #5) with cost =
$3.0/sec
11.00 Sets the submission time to 11.00 ****
11.00 Sets Gridlet status from Created to InExec
11.00 Sets the execution start time to 11.00
112.00 Sets Gridlet status from InExec to Success
112.00 Sets the wall clock time to 101.00 and the actual CPU time to
101.00
112.00 Sets the length's finished so far to 1000.0
Gridlet #0, length = 1000.0, finished so far = 1000.0
===========================================
-------End of QUOTE----------------------
------QUOTE - User1 - Gridlet 1 result -------
Time (sec) Description Gridlet #0
------------------------------------------
0.00 Creates Gridlet ID #0
0.00 Assigns the Gridlet to User_1 (ID #20)
19.00 Allocates this Gridlet to GridResource_0 (ID #5) with cost =
$3.0/sec
19.00 Sets the submission time to 19.00 ****
19.00 Sets Gridlet status from Created to InExec
19.00 Sets the execution start time to 19.00
120.00 Sets Gridlet status from InExec to Success
120.00 Sets the wall clock time to 101.00 and the actual CPU time to
101.00
120.00 Sets the length's finished so far to 1000.0
Gridlet #0, length = 1000.0, finished so far = 1000.0
===========================================
-------End of QUOTE----------------------
Because, for these 2 gridlets, both are start immediately once the it is
ready for execution. So, now, let's put topic of the execution time off
for a second. From the line with highlight, I doubt that even User1 has
5x higher bandwidth. Why it is slower (ready at 19th sec while 11th sec
for User0)? And if you see from my previous email, this happens for
every pairs of gridlet. So, what is the factor that make it "ready for
submission to PE" later than User0?
Sorry for very long email. I think it starts to sound complicate to me.
Please kindly let me know, if further information is needed. Thank you
for your valuable time in helping this.
Best Regards,
Chainan Satayapiwat
Research Student, Tohoku University
On Thu, 2007-12-27 at 10:11 +0100, Agustín Caminero Herráez wrote:
> Hi Chainan,
>
> I think I can give some hints to fix your problem.
>
> First, you say that "- Gridlet#0 size 1000 submitted at 11.00 starts at
> 11.00 ". This means, to my understanding, that the gridlet is being
> submitted at time 11.00 and its execution starts at 11.00. This means
> that the second 11.00 includes the network transmission time + the time
> to wait for a idle PE at the resource. Since this is the first gridlet,
> all PEs are idle so it starts being executed straight-away.
> If u pay attention, the start time is the same as the submission time
> for the first 3 gridlets in your sims, because your resource has 3 PEs.
> Delays begin when all the PEs are busy, so they are created because of
> congestion at the resource.
>
> In order to test the network performance, you can create a bigger
> resource (with more PEs) so that time differences between gridlets will
> be caused by the network transmission, not by contention at the
> resource. If there are no differences, try making the gridlets IO files
> bigger.
>
> Hope this is useful.
>
> All the best,
>
> Agustin
>
> Chainan Satayapiwat escribio':
> > Hi,
> >
> > I was studying the sample code which provided in
> > http://www.gridbus.org/gridsim/example/index.html Specifically, my
> > problem was coming from the example 7 (Test.java & TestCase1.java) when
> > running the simulation. The allocation policy was space sharing.
> >
> > There are 2 users (User_0 and User_1) I saw all the Gridlet from User_0
> > start queuing for processing earlier than User_1 but the lenght of
> > Gridlet (which seems to use as Gridlet size, in_file size, and out_file
> > size) are all the same. Given the User_1 has 5 times higher bandwidth,
> > then, I would assume that the User_1 start time should be earlier than
> > User_0. However, here are the result of simulation
> >
> > Here are the detail.
> > All Gridlet schedules onto 1 Machine with 3 PEs.
> > User_0 (Bandwidth 1000) has
> > - Gridlet#0 size 1000 submitted at 11.00 starts at 11.00
> > - Gridlet#1 size 2000 submitted at 35.00 starts at 35.00
> > - Gridlet#2 size 3000 submitted at 75.00 starts at 120.00
> > - Gridlet#3 size 4000 submitted at 131.00 starts at 313.00
> >
> > User_1 (Bandwidth 5000) has
> > - Gridlet#0 size 1000 submitted at 19.00 starts at 19.00
> > - Gridlet#1 size 2000 submitted at 51.00 starts at 112.00
> > - Gridlet#2 size 3000 submitted at 99.00 starts at 236.00
> > - Gridlet#3 size 4000 submitted at 163.00 starts at 421.00
> >
> > As shown above, I can't understand how it works here. User with higher
> > bandwidth should be able to finish uploading the file to the processing
> > machine first and, thus, the job should be ready to submit earlier.
> > Therefore, I tried to make sure if this is about bandwidth things by
> > hardcode in the TestCase1.java in createGridlet method by fixing the
> > size of input file. (i.e. third parameter of Gridlet constructor) As
> > expected, all Gridlet starts at 3rd second of simulation time. (This is
> > considered immediate as the code wait 3 units of time in the beginning)
> > So, I'm quite sure that the delay is coming from the simulation of file
> > transfer that required before the program execution. So, why the higher
> > bandwidth is slower? Perhaps, I misunderstood at some points and confuse
> > myself. Could anyone kindly help me out on this?
> >
> > Thank you.
> >
> > Best Regards,
> > Chainan Satayapiwat
> > Research Student, Tohoku University
> >
> >
> > -------------------------------------------------------------------------
> > This SF.net email is sponsored by: Microsoft
> > Defy all challenges. Microsoft(R) Visual Studio 2005.
> > http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/
> > _______________________________________________
> > Gridsim-users mailing list
> > Gri...@li...
> > https://lists.sourceforge.net/lists/listinfo/gridsim-users
> >
>
|
