Replacement policy

From gem5
Revision as of 05:03, 1 October 2019 by Odanrc (talk | contribs) (Clarify invalidation behaviour)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to: navigation, search

Gem5 has multiple implemented replacement policies. Each one uses its specific replacement data to determine a replacement victim on evictions.

All of the replacement policies prioritize victimizing invalid blocks.

A replacement policy consists of a reset(), touch(), invalidate() and getVictim() methods. Each of which handles the replacement data differently.

  • reset() is used to initialize a replacement data (i.e., validate). It should be called only on entry insertion, and must not be called again until invalidation. The first touch to an entry must always be a reset().
  • touch() is used on accesses to the replacement data, and as such should be called on entry accesses. It updates the replacement data.
  • invalidate() is called whenever an entry is invalidated, possibly due to coherence handling. It makes the entry as likely to be evicted as possible on the next victim search. An entry does not need to be invalidated before a reset() is done. When the simulation starts all entries are invalid.
  • getVictim() is called when there is a miss, and an eviction must be done. It searches among all replacement candidates for an entry with the worst replacement data, generally prioritizing the eviction of invalid entries.

We briefly describe the replacement policies implemented in Gem5. If further information is required, the Cache Replacement Policies Wikipedia page, or the respective papers can be studied.

Random

The simplest replacement policy; it does not need replacement data, as it randomly selects a victim among the candidates.

Least Recently Used (LRU)

Its replacement data consists of a last touch timestamp, and the victim is chosen based on it: the oldest it is, the more likely its respective entry is to be victimized.

Tree Pseudo Least Recently Used (TreePLRU)

A variation of the LRU that uses a binary tree to keep track of the recency of use of the entries through 1-bit pointers.

Bimodal Insertion Policy (BIP)

The Bimodal Insertion Policy is similar to the LRU, however, blocks have a probability of being inserted as the MRU, according to a bimodal throttle parameter (btp). The highest btp is, the highest is the likelihood of a new block being inserted as MRU.

LRU Insertion Policy (LIP)

The LRU Insertion Policy consists of a LRU replacement policy that instead of inserting blocks with the most recent last touch timestamp, it inserts them as the LRU entry. On subsequent touches to the block, its timestamp is updated to be the MRU, as in LRU. It can also be seen as a BIP where the likelihood of inserting a new block as the most recently used is 0%.

Most Recently Used (MRU)

The Most Recently Used policy chooses replacement victims by their recency, however, as opposed to LRU, the newest the entry is, the more likely it is to be victimized.

Least Frequently Used (LFU)

The victim is chosen using the reference frequency. The least referenced entry is chosen to be evicted, regardless of the amount of times it has been touched, or how long has passed since its last touch.

First-In, First-Out (FIFO)

The victim is chosen using the insertion timestamp. If no invalid entries exist, the oldest one is victimized, regardless of the amount of times it has been touched.

Second-Chance

The Second-Chance replacement policy is similar to FIFO, however entries are given a second chance before being victimized. If an entry would have been the next to be victimized, but its second chance bit is set, this bit is cleared, and the entry is re-inserted at the end of the FIFO. Following a miss, an entry is inserted with its second chance bit cleared.

Not Recently Used (NRU)

Not Recently Used (NRU) is an approximation of LRU that uses a single bit to determine if a block is going to be re-referenced in the near or distant future. If the bit is 1, it is likely to not be referenced soon, so it is chosen as the replacement victim. When a block is victimized, all its co-replacement candidates have their re-reference bit incremented.

Re-Reference Interval Prediction (RRIP)

Re-Reference Interval Prediction (RRIP) is an extension of NRU that uses a re-reference prediction value to determine if blocks are going to be re-used in the near future or not. The higher the value of the RRPV, the more distant the block is from its next access. From the original paper, this implementation of RRIP is also called Static RRIP (SRRIP), as it always inserts blocks with the same RRPV.

Bimodal Re-Reference Interval Prediction (BRRIP)

Bimodal Re-Reference Interval Prediction (BRRIP) is an extension of RRIP that has a probability of not inserting blocks as the LRU, as in the Bimodal Insertion Policy. This probability is controlled by the bimodal throtle parameter (btp).