Thanks Andreas. This makes a dramatic difference in the results.

Regards,


On 2 June 2014 15:13, Andreas Hansson <[email protected]> wrote:

>  Hi Ahmad,
>
>  You can tune the page policies of both. By default the DRAMCtrl model in
> gem5 uses an open-adaptive policy with FR-FCFS scheduling, which tends to
> be the best possible when it comes to row hits. You can change this if you
> want, just modify the page_policy parameter of the controller instance. The
> default in DRAMSim2 is a simpler open-page policy with a limit of 4 access
> to the same page before closing it. Thus, you are far more likely to get
> hit by the opening/closing.
>
>  I hope that all makes sense.
>
>  Andreas
>
>
>   From: Ahmad Hassan via gem5-users <[email protected]>
> Reply-To: Ahmad Hassan <[email protected]>, gem5 users mailing list <
> [email protected]>
> Date: Monday, 2 June 2014 15:04
>
> To: gem5 users mailing list <[email protected]>
> Subject: Re: [gem5-users] DDR timing model
>
>  Hi Andreas,
>
> Thanks for the details. But the same workload gives significant increase
> in average memory access time if i run gem5 with integrated dramsim2 and
> increase only tRCD and tRP. As you said, both tRCD and tRP are associated
> with opening and closing of row. The results of dramsim2 and gem5 DDR3
> memory model suggests that gem5 keeps the rows open for much longer
> duration which causes significantly higher pageHitRate.
>
>  Regards,
>
> On 2 June 2014 10:36, Andreas Hansson <[email protected]> wrote:
>
>>  Hi Ahmad,
>>
>>  Increasing tRCD and tRP will essentially make it more costly to open
>> and close a page. If you have good page hit-rate, and good distribution
>> over the banks, then the effects of the increased timings can be hidden (at
>> least from the perspective of the average latency). When you increase tCL
>> you are delaying every single access.
>>
>>  bytesPerActivate is a histogram statistic associated with the DRAM
>> controller. You should find it in the stats.txt along with all the other
>> stats.
>>
>>  Andreas
>>
>>   From: Ahmad Hassan via gem5-users <[email protected]>
>> Reply-To: Ahmad Hassan <[email protected]>, gem5 users mailing list
>> <[email protected]>
>> Date: Monday, 2 June 2014 10:29
>> To: gem5 users mailing list <[email protected]>
>> Subject: [gem5-users] DDR timing model
>>
>>    Hi Andreas,
>>
>> [Moving discussion from http://reviews.gem5.org/r/2109]
>>
>>  If I only increase the tCL parameter from 13.75ns to 61ns [4.4x times
>> default DDR3), then I notice 3.2x times increase in avgMemAccLat. This is
>> what we should expect. Can we use tCL to simulate slow DDR3?  My
>> understanding is that, we cannot do that because tCL represents column
>> address strobe and it affects bandwidth.
>>
>>  But I cannot produce the same result by changing tRCD = '61ns', tRP =
>> '165ns'. tRCD affects read latency and tRP affects write latency. I only
>> get an increase in avgMemAccLat by 1.2x times default DDR3.
>>
>>  Results:
>>
>> For default DDR3:
>> avgMemAccLat                22133.24
>> totMemAccLat             12166996811
>>
>> For tRCD = '61ns', tRP = '165ns'.
>> avgMemAccLat                27473.08
>> totMemAccLat             15104591550
>>
>> For tCL = '61ns'
>> avgMemAccLat                71630.94
>> totMemAccLat             39376746052
>>
>>  I did not find any 'bytesPerActive' attribute in stats file. Where shall
>> I see this?
>>
>>  Thanks.
>>
>>
>>
>>
>>
>> -- IMPORTANT NOTICE: The contents of this email and any attachments are
>> confidential and may also be privileged. If you are not the intended
>> recipient, please notify the sender immediately and do not disclose the
>> contents to any other person, use it for any purpose, or store or copy the
>> information in any medium. Thank you.
>>
>> ARM Limited, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ,
>> Registered in England & Wales, Company No: 2557590
>> ARM Holdings plc, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ,
>> Registered in England & Wales, Company No: 2548782
>>
>
>
> -- IMPORTANT NOTICE: The contents of this email and any attachments are
> confidential and may also be privileged. If you are not the intended
> recipient, please notify the sender immediately and do not disclose the
> contents to any other person, use it for any purpose, or store or copy the
> information in any medium. Thank you.
>
> ARM Limited, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ,
> Registered in England & Wales, Company No: 2557590
> ARM Holdings plc, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ,
> Registered in England & Wales, Company No: 2548782
>
_______________________________________________
gem5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users

Reply via email to