Hi Prathap,

I assume the core are sharing an L2? How recent is your version of gem5? The 
latency of the snoops was only recently accounted for.

In cases like this it also helps to actually trace a few memory accesses with 
the debug flags (Cache and Xbar in this case).

Andreas

From: gem5-users 
<[email protected]<mailto:[email protected]>> on behalf of 
Prathap Kolakkampadath <[email protected]<mailto:[email protected]>>
Reply-To: gem5 users mailing list 
<[email protected]<mailto:[email protected]>>
Date: Tuesday, 8 December 2015 at 17:06
To: gem5 users mailing list <[email protected]<mailto:[email protected]>>
Subject: [gem5-users] Snoop Timing

Hello Users,

Simulation Setup
--------------------------
Full System Mode, Classic memory system, Four O3(arm_detailed)  cpus

Test Setup (solo)
--------------------
- Running  a Read thread(software thread)
- The thread allocates a buffer that fits in L1 Cache(<32KB).
-  Loops over the buffer to generate a fixed amount of L1 cache Reads.
-  All are L1 Read Hits.
- Observed the time taken to complete the thread.

Test Setup  (3 interfering threads)
-----------------------------------------------
- Repeated the above experiment with three interfering Write threads on other 
three cores.
- The interfering threads shares the same memory with the Test Under 
Analysis(TuA) and continuously writes to the shared memory.

With three interfering write threads, now all the L1 read access of TuA, now 
misses the L1. This is expected as write invalidates the cache lines of all 
other cores. Now the read misses needs to snoop from other core private caches. 
Bu this should incur some delay.
However I see no slowdown. I believe snooping from other private cache has to 
pay more latency than the L1 hit latency.  I am not sure if this is modelled in 
Gem5. Can someone please confirm?

Thanks,
Prathap

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.
_______________________________________________
gem5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users

Reply via email to