[ndnSIM] Inconsistency of Random replacement policy

Sun Jun 9 20:15:46 PDT 2013

Dear Saran,

I think if you try with different seeds and get the average, the difference
may be disappeared even for s=1.0.

Best Regards,
Saeid

On Fri, Jun 7, 2013 at 11:06 PM, Saran Tarnoi <sarantarnoi at gmail.com> wrote:

> Dear Saeid and All,
>
> The exact number of the content objects in the network is 10,000 different
> objects (one interest for one content object.)
> The Zipf parameters q = 0 and s = 0.5, 1.0, 1.5, and 2.0. In other words,
> Zipf-Mandelbrot is simplified to Zipf distribution.
> There is no link bottleneck due to the redundant link capacity, so all the
> interests are satisfied.
>
> The differences of hit rate depend on Zipf parameters.
> The summary of the cache hit percentage of node 1, which uses Random as
> replacement policy (the replacement policy of node 2 is varied), is below.
>
>   s      Lfu          Lru        Random
> 0.5    1.865      1.865       2.572
> 1.0    33.123    33.123     36.467
> 1.5    86.542    86.542     86.025
> 2.0    98.646    98.646     98.348
>
> The difference is obvious when s=1.0.
> If you need more information, please let me know.
> Thank you.
>
> Regards,
> Saran
>
>
>
> 2013/6/7 Saeid Montazeri <saeid.montazeri at gmail.com>
>
>> Dear Saran,
>>
>> 1. May we know the total number of contents in the network. In addition,
>> what is the size distribution of the contents. (in total how many different
>> Data packets exist in the network).
>>  2. What is the difference between cache hit ratios?
>>
>> Best Regards,
>> Saeid
>>
>>
>> On Fri, Jun 7, 2013 at 7:52 PM, Saran Tarnoi <sarantarnoi at gmail.com>wrote:
>>
>>> Dear Saeid and All,
>>>
>>> Thank you very much for your helpful answer. :)
>>> I also agree with your first thought. Two nodes may use the same
>>> generated sequence of random seed.
>>>
>>> In my experiment, the number of generated interests was quite small,
>>> 200,000 interests with 60,000 interest warm up period.
>>> The cache size is 100 objects.
>>> The scale is tiny but the difference of results obviously appears.
>>>
>>> I'm running simulation for 2,000,000 interests with 1,000,000 interest
>>> warm up period to see whether or not the interest number can help.
>>> Thank you again.
>>>
>>> Regards,
>>> Saran Tarnoi
>>>
>>>
>>>
>>> 2013/6/7 Saeid Montazeri <saeid.montazeri at gmail.com>
>>>
>>>> Dear Saran,
>>>>
>>>> Would you please let us more about your experiment setting: how long
>>>> the consumer generates Interests or how many requests are generated?
>>>>
>>>> Meanwhile, my guess is that the problem may be because of the sequence
>>>> of the random number that is used by random replacement policies. *If
>>>> both policies in node 1 and 2 use the same random number generator(or at
>>>> least dependent random number generators)*, the problem is because of
>>>> that. If we assume that is the case, suppose that a random number sequences
>>>> for cache size equals 2 be 1,2,1,2,2,1. (3 means that incoming Data packet
>>>> will not be written to the cache). If only one replacement policy is random
>>>> then 1,2,3,1,3,2,2,1 will be the replaced slots in the first node. However,
>>>> if both of the nodes have random replacement policies, the replaced slot in
>>>> the first node depends on the misses in the second node. *If the
>>>> policies in node 1 and 2 use independent random generators*, I have no
>>>> idea.
>>>> In any case, would you please let us know for how many requests you did
>>>> the experiment. Because I think it should be related to the
>>>> transient behavior of the system and if you run the simulation for enough
>>>> long period of time, you may get pretty close results.
>>>>
>>>> Best Regards,
>>>> Saeid
>>>>
>>>>
>>>>
>>>> On Fri, Jun 7, 2013 at 4:19 PM, Saran Tarnoi <sarantarnoi at gmail.com>wrote:
>>>>
>>>>> Hello Alex and All,
>>>>>
>>>>> I did some experiments and found some wired results.
>>>>> The experimental setup is as follows.
>>>>>
>>>>> (Random, Lru, or Lfu)   node 2 <--- (producer)
>>>>>                                      l
>>>>>                                      l
>>>>>         (Random)            node 1 <--- (consumer, Zipfmandelbrot)
>>>>>
>>>>> The cascading network has 2 nodes.
>>>>> These nodes has the same size of Content Store.
>>>>> Consumer and producer are installed in nodes 1 and 2, respectively.
>>>>> Zipfmendelbrot is set as the Interest traffic model for consumer.
>>>>>
>>>>> Random is the replacement policy for node 1 while either Random, Lru,
>>>>> or Lfu is the replacement policy of node 2.
>>>>>
>>>>> The cache hit rate at node 1 (the first-hop node) should be consistent
>>>>> and independent of the replacement policy deployed in node 2.
>>>>> However, the results don't follow the above logic.
>>>>>
>>>>> The cache hit rate at node 1 when either Lru or Lfu is set as the
>>>>> replacement policy of node 2 are exactly the same but different from when
>>>>> Random is the replacement policy of node 2.
>>>>>
>>>>> In addition, I don't see any problem when Lfu or Lru is set as
>>>>> replacement policy of node 1.
>>>>>
>>>>> Is it a bug or random seed problem?
>>>>> Could you give me any idea?
>>>>> Thanks a lot for your time.
>>>>>
>>>>> Regards,
>>>>> Saran Tarnoi
>>>>>
>>>>> _______________________________________________
>>>>> ndnSIM mailing list
>>>>> ndnSIM at lists.cs.ucla.edu
>>>>> http://www.lists.cs.ucla.edu/mailman/listinfo/ndnsim
>>>>>
>>>>>
>>>>
>>>
>>>
>>> --
>>> Regards,
>>> Saran Tarnoi
>>> Graduate Student
>>> Department of Informatics
>>> The Graduate University for Advanced Studies (Sokendai)
>>> Tokyo, Japan
>>> http://sarantarnoi.blogspot.jp<http://sarantarnoi.blogspot.jp/2012/11/my-profile.html>
>>>
>>
>>
>
>
> --
> Regards,
> Saran Tarnoi
> Graduate Student
> Department of Informatics
> The Graduate University for Advanced Studies (Sokendai)
> Tokyo, Japan
> http://sarantarnoi.blogspot.jp<http://sarantarnoi.blogspot.jp/2012/11/my-profile.html>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.lists.cs.ucla.edu/pipermail/ndnsim/attachments/20130610/2418d1f3/attachment.html>