[Question] How do you write a Scarab data prefetcher? about scarab HOT 2 CLOSED

I'm not particularly familiar with the prefetcher codebase, but I took a look and can answer your questions at a high level and point you towards the right direction.

A prefetcher in Scarab is defined using the HWP struct (or HWP_struct). HWP contains all relevant data/metadata for a prefetcher and a bunch of function pointers. HWP is essentially a class interface implemented in C using function pointers. To add a new prefetcher, you need to define all the relevant functions for a new HWP struct. These functions are then called for all prefetchers at the corresponding event. Then, you'll add an entry to the array of HWPs as defined in src/prefetcher/pref_table.def. Note that you do not have have to define all the functions, if you look at the currently implemented prefetchers in, you can see that many of the function pointers are NULL for various prefetchers.

pref_stream_ul1_miss is the function corresponding to ul1_miss_func in the HWP of the stream prefetcher. To be able to figure out what it does, you can look up where ul1_miss_func is called in the code. It is called by pref_ul1_miss() in pref_common.c, which is called from mem_complete_l1_access() in memory.c when an access misses in the L1 cache.

Side note: in Scarab, for historical reasons, L1 cache refers to the LLC.

It seems like pref_stream_throttle and pref_stream_train_stream_filter implement a mechanism for throttling how much the stream prefetcher consumes bandwidth. You need to dig into the code to understand how it works. Note that there is some amount of dead code in Scarab, so if you want to go through some code, I recommend assuming Scarab is configured using src/PARAMS.kaby_lake on top of the defaut parameters.

from scarab.

Closing this due to inactivity. Feel free to reopen if there are followup questions.

from scarab.