[Gems-users] 回复： 回复： 回复： 回复： 回复： 回复： Latency and number of memory request

["张轶" <zhangyi@xxxxxxxxxxxxxx>]

Dan,

 

Many thanks for your help that I could basically understand the internal mechanism of GEMS's memory controller. I am sorry I still have some questions about memory channel's configuration.

 

In GEMS default setting, each chip would have one memory channel(memory controller). I would like to implement such a config that the number of memory channel is less than the chip number. Also, I would add some arbitrations between channels.

 

In your past mail, you told me I should look at GEMS's topology. But so sorry, by far i haven't understood it well. Would you please tell me some further details about configuring the memory channel? Where should i start up in topology configuration?

 

Regards,

Zhang Yi

 

------------------ 原始邮件 ------------------

发件人: "Dan Gibson"<degibson@xxxxxxxx>;

发送时间: 2010年5月14日(星期五) 晚上9:15

收件人: "Gems Users"<gems-users@xxxxxxxxxxx>;

主题: Re: [Gems-users]回复： 回复： 回复： 回复： 回复： Latency and number of memory request

 

Memory controller internal delay models delay from logic pathways in the memory controller that don't fall into other categories. I.e., if you want to make the memory pipeline a little longer, add cycles there.

Regards,

Dan

2010/5/13 张轶 <zhangyi@xxxxxxxxxxxxxx>

Hi Dan,

Thanks for your help. My understanding is not correct.

 

Btw, in the MEM_CTL_LATENCY, does the "(four bit times) + (round trip on channel) " represent the time costed on data bus? And which delay does "memory controller internal delay" represent?

 

Regards,

Zhang Yi

 

 

------------------ 原始邮件 ------------------

发件人: "Dan Gibson"<degibson@xxxxxxxx>;

发送时间: 2010年5月12日(星期三) 晚上9:12

收件人: "Gems Users"<gems-users@xxxxxxxxxxx>;

主题: Re: [Gems-users]回复： 回复： 回复： 回复： Latency and number of memory request

 

Zhang,

According to my foggy recollection of the GEMS memory controller, MEM_CTL_LATENCY ends up as part of the enqueue-message delay, and BANK_BUSY_TIME is part of modeling the utilization of the banks themselves. E.g.:

At time T, a request is sent to Bank B. Total service time for that request is MEM_CTL_LATENCY + MEM_FIXED_DELAY, so when the request is issued, a response message is enqueued with that delay (the delay of the memory is modeled as a queueing delay in MemoryController.C). However, Bank B will be in operation servicing the request for BANK_BUSY_TIME, which is less than MEM_CTL_LATENCY, so the bank is marked as busy that number of cycles as well.

In other words, it is not double-charging for tRCT+CL+AL -- it is simply accounting for delay and contention.

Regards,

Dan

2010/5/12 张轶 <zhangyi@xxxxxxxxxxxxxx>

Hi Dan,

I am wondering if the default setting of request's latency in memory controller is wrong?

 

In the rubyconfig.default, there is the parameter "MEM_CTL_LATENCY: 12" with the comment "This equals tRCD + CL + AL + (four bit times) + (round trip on channel) + (memory control internal delays)".

And the BANK_BUSY_TIME is "Bank cycle time (tRC) measured in memory cycles".

 

In the memorycontroll.c, there is MemoryControl::issueRequest{

...

if (req.m_msgptr.ref() != NULL) {
    enqueueToDirectory(req, m_mem_ctl_latency + m_memFixedDelay);
  }

...

} and the m_mem_ctl_latency = MEM_CTL_LATENCY

 

So I think the MEM_CTL_LATENCY has been calculated in the memory request latency. Meanwhile in a request routine, the BANK_BUSY_TIME is also calculated in a memory request.

But in memory system, tRC includes tRCD + CL + AL .

 

According to those information, I am wondering if the latency of a request costed on memory controller has been calculated in additional time "tRCD + CL + AL "?

 

Regards,

Zhang Yi

 

 

--

http://www.cs.wisc.edu/~gibson [esc]:wq!

_______________________________________________
Gems-users mailing list
Gems-users@xxxxxxxxxxx
https://lists.cs.wisc.edu/mailman/listinfo/gems-users
Use Google to search the GEMS Users mailing list by adding "site:https://lists.cs.wisc.edu/archive/gems-users/" to your search.

--
http://www.cs.wisc.edu/~gibson [esc]:wq!

_______________________________________________
Gems-users mailing list
Gems-users@xxxxxxxxxxx
https://lists.cs.wisc.edu/mailman/listinfo/gems-users
Use Google to search the GEMS Users mailing list by adding "site:https://lists.cs.wisc.edu/archive/gems-users/" to your search.

--
http://www.cs.wisc.edu/~gibson [esc]:wq!