[Disksim-users] DiskSim 3.0/FlashSim Simulation problems

Fri Mar 25 16:55:08 EDT 2011

That's exactly why a trace should be submitted to your SSD so that a new request is scheduled only when the previous one has been completed (this is what I mean by closed subsystem model). In this way you are able to evaluate the performance of various FTL schemes given a same access sequence (trace)
In other words your original trace is intensive just because, being the requests dispatched in parallel over more than one SSD, the average execution time is smaller causing next access arrival time to be nearer to the previous.

Does anyone know hot to configure the simulator so that even with an external trace a new request is scheduled only when the previous one has been executed (making simulator ignore the 'arrival times') ?

Regarding disksim I am just trying to run a debugging session with code block to get more familiar with the code but I am not so familiar with Linux, makefile. I am having trouble when giving 'make clean' and 'make'.
It seems that 'clean' command changes also all the .c files causing the original compilation error :(.

What IDE are you using?

Francesco
  From: Jonathan Tjioe 
  To: Francesco Falanga 
  Cc: disksim-users at ece.cmu.edu 
  Sent: Thursday, March 24, 2011 1:32 AM
  Subject: Re: [Disksim-users] DiskSim 3.0/FlashSim Simulation problems

  Francesco,

  I'm glad my previous procedures were helpful to you. That's why I wrote them =)

  I understand what you are saying about changing the subsystem. For example, I am simulating only 1 SSD. The original Financial traces are simulating multiple hard disks. I have modified the trace so that for devicenum, it is always 0 (hard disk 0). I believe this is a valid modification to the trace since it doesn't not change the request arrival times, size of each request, etc. It is only changing the physical disks I have available to run the simulation through. However, I'm not sure that changing the arrival times is a legitimate change. For example, the Financial traces are very write intensive - the time between writes is very small, i.e. the write request rate is very high. If I changed these arrival times to be further spaced apart, I cannot legitimately say that I ran my GC algorithm or FTL through the Financial trace as now it is not as write intensive as it should be. However, changing things about my subsystem: how many disks I have, what kind of disks they are, etc. are completely valid changes. Hopefully that clears up what I was trying to say.

  With that said, I'm not sure if I understand exactly what you meant by this:
  <<
  Regarding this I am still in trouble since I read somewhere that it is possible to configure simulator in order to make it behave as a closed subsystem model also on external I/O traces making it ignore the 'arrival times' have you an idea on how this can be done?
  >>

  I'm not sure how that can be done.

  I read the link you sent me. I don't think I am implementing any magnetic disks. Do you know where I can look in the code to check exactly which disks are implemented? (I'm not that familiar with DiskSim yet).

  Thanks,
  Jonathan

  On Wed, Mar 23, 2011 at 12:44 AM, Francesco Falanga <ffalanga at fastwebnet.it> wrote:

    Hi Jonatan,

    I just find out that I am following your same steps with FlashSim encouneing problem that you've already solved :)...
    Regarding your statement:

    "So if I slowed down the request-ratio, this would mean modifying the arrival times, which would invalidate the trace as it would no longer be real world"

    I have to respectfully disagree since a "trace" gives you the performance of a very specific system configuration. Changing one or more subsystem may have still sense in order to see how the new configuration will perform given the same request sequence.

    Regarding this I am still in trouble since I read somewhere that it is possible to configure simulator in order to make it behave as a closed subsystem model also on external I/O traces making it ignore the 'arrival times' have you an idea on how this can be done?

    Regarding segmentation fault I found this, but I guess you have already given a glance:
    https://sos.ece.cmu.edu/pipermail/disksim-users/2010-December/000556.html

    Ciao!

    Francesco.

      ----- Original Message ----- 
      From: Jonathan Tjioe 
      To: Alexander Lochmann 
      Cc: Aayush Gupta ; Andres Blanco ; disksim-users at ece.cmu.edu ; Youngjae Kim 
      Sent: Wednesday, March 23, 2011 3:47 AM
      Subject: Re: [Disksim-users] DiskSim 3.0/FlashSim Simulation problems

      Alex,

      I've moved your responses up to the top so the rest of the forum won't get confused trying to distinguish the your responses from my questions.

      I've also CCed Aayush Gupta and Youngjae Kim as they were the authors of the DFTL paper that my environment is based on. Hopefully, they can shed some light on what I'm doing wrong.

      On your response to Error #1 (Segmentation fault)...

      When I run it for the Financial1 trace file using all 3 FTLs, I get the following output to screen:

      Running Pagemap FTL...
      start_blk_no: 1107318608, block_cnt: 4, total_util_sect_num: 2097152
      Running DFTL...
      ./runtest: line 11:  5905 Segmentation fault      ../src/disksim dftl.parv dftl.outv ascii ./trace/test.file 0
      Running FAST...
      start_blk_no: 1107318608, block_cnt: 4, total_util_sect_num: 2097152

      As stated earlier, if I check the .outv files that were generated for each of these FTLs, none of the simulations ever completed.

      I have not run it using gdb. The code is compiled already. I suppose I could try to do so, but I would think that I should be able to run it as is.

      I have also attached the exact script that I run when I simulate the 3 FTLs. Again, this is the same exact environment that the DFTL authors used.

      On your response to Error #2 (simulation stopped due to saturation), when running runtest script against the Financial2 trace file, I get this output to the screen.

      Running Pagemap FTL...
      Stopping simulation because of saturation: simtime 108.435867, totalreqs 10394
      IOdriver Response time average:         17.003924
      IOdriver Response time std.dev.:        15.487314
      Running DFTL...
      Stopping simulation because of saturation: simtime 107.151359, totalreqs 10255
      IOdriver Response time average:         25.488591
      IOdriver Response time std.dev.:        19.068197
      Running FAST...
      Stopping simulation because of saturation: simtime 106.554985, totalreqs 10188
      IOdriver Response time average:         10.128997
      IOdriver Response time std.dev.:        8.028486

      As far as slowing down the request-ratio...well, I could be misunderstanding what you are saying, but I thought that the way it works is that the real world requests come in exactly as the arrival times state in the trace file. Then if the hard drive(s) is(are) busy, then they just go in the request queue waiting to be serviced. I could understand that for a synthetically generated trace, it might be worth slowing down the request ratio just to see what happens, but this is a real world trace. So if I slowed down the request-ratio, this would mean modifying the arrival times, which would invalidate the trace as it would no longer be real world.

      Thanks,
      Jonathan

      On Tue, Mar 22, 2011 at 1:09 AM, Alexander Lochmann <alexander.lochmann at tu-dortmund.de> wrote: 
      Hi!

      On Error #1 (Segmentation fault), could give us some more information?
      Have you tried to run it with gdb und typed "bt"?

      On Error #2 (Simulation stopped due to saturation), maybe your harddisk model is too slow to serve the requests within a appropriate amount of time so the requestqueue doesn't get saturated. In a real system you've got a operating system which keeps track of this issue. Linux for example, slows down every readahead and writeback activity to reduce the number of requests if it detects congestion conditions. It has a requestqueue which is on top of the driver holding every request.
      Have you tried to slow down the request-ratio?

      Greetings
      Alex

        Am 22.03.2011 03:05, schrieb Jonathan Tjioe: 
          DiskSim users,

          First, let me apologize for making this email so long. I wanted it to be thorough so my problem is explained clearly.

          I'm running DiskSim 3.0 with FlashSim (same as the FlashSim in the DFTL paper). It's basically DiskSim 3.0 with support for SSD.

          I've been running the "runtest" script which basically runs a small sample test file through all 3 FTLs: Page mapped, DFTL, and FAST. The test trace that was included is very small (8-9MB) and I get the intended results as the authors of the DFTL paper did. 

          All the runtest script does is run the same trace file on these 3 FTLs:
          ../src/disksim pagemap.parv pagemap.outv ascii ./trace/test.file 0
          ../src/disksim dftl.parv dftl.outv ascii ./trace/test.file 0
          ../src/disksim fast.parv fast.outv ascii ./trace/test.file 0

          Upon successful completion of the simulation, I noticed that in the .outv files, I will see:
          <<
          ...
          loadparams complete
          Initialization complete
          Simulation complete
          ...
          >>

          After this, I will see many performance statistics along with every request. I get the appropriate results when simulating DFTL, FAST, and pure page mapped FTL.

          However, once I tried putting any other real world trace (which is much longer) in the simulation, it does not complete successfully. It should be noted that I have not made any modifications to any of the 3 FTLs during these tests. I have verified that my format of the trace files is correct:
          <arrival time> <devicenum> <LBA> <size in # of sectors> <1 or 0 for read or write>

          There are 3 different results that I get, all of which are unsuccessful simulations:
          1) Segmentation fault
          2) Simulation is stopped because of saturation 
          3) Simulation seems like it finished, but when you check the .outv files, the last line says "initialization complete" (in otherwords it does not ever say "simulation complete"

          One example (Error #1) is when I run the entire finanical trace in, I get a segmentation fault simulating DFTL. Although I didn't get segmentation faults for FAST or pure page mapped, their corresponding .outv files do not look correct. They seem like they never finished simulating. In the .outv files for all 3 FTLs, it shows:

          <<
          ...
          loadparams complete
          Initialization complete
          >>

          But it never shows "simulation complete" and thus the results are not shown in the .outv file. The weird thing is that if I just include maybe several hundred lines of the Financial1 trace instead of the entire thing, it completes successfully with no problem, so I know my format of the file is correct.

          Another example (Error #2) I noticed that I the Financial2 trace did not have any segmentation faults but instead I had a different message:
          <<
          Stopping simulation because of saturation: simtime 108.435867, totalreqs 10394
          >>

          I did some research and found that there is a define statement in disksim_logorg.c that sets the MAX_QUEUE_LENGTH to be 10000. In each of the instances when the simulation stopped due to saturation, the totalreqs number was just over 10000. I imagine the queue length is getting very large due to the fact that I only have a device num of 0 since my GC algorithm will be a local GC algorithm, not a inter-disk algorithm.

          And lastly, if (Error #1) or (Error #2) occur, the simulation will never complete and no simulation summary will be in the .outv file (which is Error #3).

          My questions are as follows:
          Regarding (Error #1), I have no idea why I am getting segmentation faults. Do you think it is some type of buffer overflow issue b/c there is so much data with the real world traces? Remember, that if I just take a smaller subset of the real world data, it simulates to completion without any problem.

          Regarding (Error #2), I can try just increasing the MAX_QUEUE_LENGTH to maybe 100000 and see what happens, but is that the right solution or is there something else that I need to be aware of.

          Regarding (Error #3), I would assume that this will be solved once I find the solutions to (Error #1) and (Error #2).

          I really appreciate any help or hints you can offer. I will also post this to the disksims mailing list.

          Thanks for your valuable time,
          Jonathan Tjioe

_______________________________________________
Disksim-users mailing list
Disksim-users at ece.cmu.edu
https://sos.ece.cmu.edu/mailman/listinfo/disksim-users

--------------------------------------------------------------------------

      _______________________________________________
      Disksim-users mailing list
      Disksim-users at ece.cmu.edu
      https://sos.ece.cmu.edu/mailman/listinfo/disksim-users

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.andrew.cmu.edu/pipermail/disksim-users/attachments/20110325/34209dad/attachment.html>