Re: [Gems-users] Interpreting cache results from ruby output

[Dan Gibson <degibson@xxxxxxxx>]

Jack,
Let me (try to) answer your questions inlined below.

Jack Tzu-Han Hung wrote:
> Hi,
>  
> Does anybody know how to interpret the results from ruby output? I 
> find it fairly hard. Below is my result from running a test program, 
> walking through an array multiple times using a loop. I use the 
> default configuration, except the L1D associativity is 1.
>  
> ------------------------------------------
> instruction_executed: 286820499 [ 286820499 ]
> cycles_per_instruction: 0.570198 [ 0.570198 ]
> misses_per_thousand_instructions: 0.258381 [ 0.258381 ]
> ...
> L1D_cache cache stats:
>   L1D_cache_total_misses: 512423
>   L1D_cache_total_demand_misses: 512423
>   L1D_cache_total_prefetches: 0
>   L1D_cache_total_sw_prefetches: 0
>   L1D_cache_total_hw_prefetches: 0
>   L1D_cache_misses_per_transaction: 512423
>   L1D_cache_misses_per_instruction: 0.00178656
>   L1D_cache_instructions_per_misses: 559.734
>
>   L1D_cache_request_type_LD:   84.6488%
>   L1D_cache_request_type_ST:   9.76849%
>   L1D_cache_request_type_ATOMIC:   5.58269%
>
>   L1D_cache_access_mode_type_SupervisorMode:   272090    53.0987%
>   L1D_cache_access_mode_type_UserMode:   240333    46.9013%
>   L1D_cache_request_size: [binsize: log2 max: 64 count: 512423 
> average: 5.94929 | standard deviation: 6.83207 | 0 14544 8970 313205 
> 169318 0 0 6386 ]
> ........
>
> Simics Driver Transaction Stats
> ----------------------------------
> Insn requests: 0
> Data requests: 115499491
> ....................................................
>
> ------------------------------------------
>  
> My questions are:
>  
> (1) What does "L1D_cache_access_mode_type_SupervisorMode" mean? I 
> suppose that's the number of cache misses (not access) from OS, but my 
> toy program doesn't have too many system calls (actually, just one 
> printf). Isn't the number too high to be reasonable?
Thats correct -- the number of supervisor cache misses. It might seem 
unreasonable, but remember that when you load ruby the caches are COLD, 
and generally the first time you run an application you page tables are 
cold, as well as TLBs and such. This is even more true of short-running 
processes like toys -- the OS starts to dominate.
>  
> (2) Since ruby can separate the misses fomr OS and user code, I wonder 
> if I can get the individual numbers of executed instructions from OS 
> and user code. Or if anyone has better ideas how we can reduce the 
> effect from OS?
Actually, Ruby doesn't go out of its way to do this -- when Simics hands 
Ruby a memory access, there is a 'priv' bit associated with that access. 
However, ruby calculates the total number of instructions by querying 
simics at the start and end of the simulation. You could augment ruby to 
count user and supervisor instructions separately, but you would need to 
add state to SimicsProcessor or SimicsDriver to do that.
>  
> (3) Does "Data requests" mean the number of toal memory accesses 
> (sending out from cpu)?
Thats correct. Data requests above is the total number of observed data 
accesses (loads, stores, and atomics) from all simics CPUs.
>  
> I really appreciate your help. Thanks a lot.
>  
> Jack
> ------------------------------------------------------------------------
>
> _______________________________________________
> 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!