Stuctures shared among cores

[igu...@cse.yeditepe.edu.tr]

Hi,

I want to isolate the impact of L3 ways assigned to a core on performance. I ran two tests:

1. A single core run where the associativity of L3 cache is 4.
   
2. A 4-core run where the associativity of L3 cache is 16. I also added a little piece of code which guarantees that the first core will use only the first 4-ways of L3 cache, and other cores won't be able to access this area. The benchmark which runs on the first core is the same one used in the single core run.

My initial expectation was to achieve the exact same performance level for the specific benchmark in these two runs. However, performance significantly drops for the benchmark in the second run, which means L3 cache is not the only structure shared among cores. Our first guess is that it is the main memory. Are there any other structures that could cause the performance difference described above; and is there an easy way to remove this difference? I am also attaching the params.in file I am using.

Thanks for any help in advance.

"Bu mesaj (ve ekleri) gizli bilgi içermektedir ve sadece gönderilen kişiye yöneliktir. Bu e-mailin muhatabı değilseniz veya içeriği ile ilginiz yoksa, Yeditepe Üniversitesinin onayı olmaksızın bu mesajın okunması, değiştirilmesi, kopyalanması, üçüncü kişilere açıklanması, yayınlanması, ifşa edilmesi veya iletilmesi yasaktır. Bu mesajın gönderilmek istendiği kişi değilseniz (ya da bu e-posta'yı yanlışlıkla aldıysanız), lütfen yollayan kişiyi hemen haberdar ediniz ve mesajı sisteminizden derhal siliniz. E-mail iletiminin güvenli veya hatasız olduğunun garantisi olmadığından geç veya eksik iletim veya içerik ve bilgilerde eksiklik, kayıp, değişiklik veya virüs olabilir. Bu nedenle, bu mesajın iletiminden dolayı, Yeditepe Üniversitesi , içerikteki hata, eksiklik, doğruluğun ve gizliliğin ihlalinden veya bu yolla bilgi paylaşımı, iletimi, depolanması gibi herhangi bir kullanımından hiçbir şekilde sorumlu değildir. Bu mesajın içeriği yazarına ait olup, Üniversitemizin görüşlerini içermeyebilir.

Bu mesajın içeriğinde geçen Üniversitemizin ad veya nanıma yaratılan fikri ve sınai haklar Üniversitemize ait olup, maddi ve manevi tüm hakları saklıdır.”

“This Message (including any attachments) contains confidential information and is intended only for the individual named. If you are not the named adressee or not related with the content of this Message, you are forbidden to read, disseminate, distribute, copy, reproduce or modify this mail by Yeditepe University. Please notfy the sender immediately if you have received this e-mail by mistake and delete this e-mail from your system. E-mail transmisson can not be guaranteed to be secure or error-free as the mail may arrive late or incomplete or the information could be intercepted, corrupted, lost, destroyed, amended, , or contain viruses. Yeditepe University therefore does not accept liability for any errors, loss of integrity or confidentiality or ommissions in the contents of this Message or for the information transmission, reception, storage of use of such in any way whatsoever, which arise as a result of e-mail transmission. Any opinions expressed in this message are those of the author and may not necessarily reflect the opinions of Our University.

Copyright in documents created by or on behalf of our University remains vested in us, and we assert all of our moral and intellectual property rights."

[Hyesoon Kim]

Hi, 

When you run 4 cores, all cores are sharing the memory system especially DRAM, so there should be performance degradations. 

If you do not want to observe memory sharing effects,  you can just modify memory.cc and have a fixed memory latency, so that there are no shared resource effects. 

Unfortunately a fixed memory latency model is not provided but it should be very easy to modify 

by returning a fixed latency from “when new_mem_req” function is called. 

Hyesoon 

--

---
You received this message because you are subscribed to the Google Groups "Macsim Developer" group.
To unsubscribe from this group and stop receiving emails from it, send an email to macsim-dev+...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.
<params.in>