System and method for partitioning of memory units into non-conflicting sets

• US 9,477,603 B2
• Filed: 09/05/2014
• Issued: 10/25/2016
• Est. Priority Date: 09/05/2013
• Status: Active Grant

• Current Assignee: Meta Platforms, Inc. (f/k/a Facebook, Inc.)
• Original Assignee: Meta Platforms, Inc. (f/k/a Facebook, Inc.)
• Inventors: Waldspurger, Carl A., Horovitz, Oded, Weis, Stephen A., Rihan, Sahil
• Primary Examiner(s): Rones, Charles
• Assistant Examiner(s): CHRZANOWSKI, MATTHEW R
• Application Number:
  US14/479,239
• Publication Number:
  US 20150067265A1
• Time in Patent Office: 781 Days
• Field of Search: None
• US Class Current: 1/1
• CPC Class Codes: G06F 12/0815 Cache consistency protocols
• CPC Class Codes: G06F 12/0862 with prefetch
• CPC Class Codes: G06F 12/0868 Data transfer between cache memory and other subsystems, e.g. storage devices or host systems
• CPC Class Codes: G06F 12/0891 using clearing, invalidating or resetting means
• CPC Class Codes: G06F 12/0895 of parts of caches, e.g. directory or tag array
• CPC Class Codes: G06F 12/126 with special data handling, e.g. priority of data or instructions, handling errors or pinning
• CPC Class Codes: G06F 2212/621 Coherency control relating to peripheral accessing, e.g. from DMA or I/O device

Abstract

A system and method of operation exploit the limited associativity of a single cache set to force observable cache evictions and discover conflicts. Loads are issued to input memory addresses, one at a time, until a cache eviction is detected. After observing a cache eviction on a load from an address, that address is added to a data structure representing the current conflict set. The cache is then flushed, and loads are issued to all addresses in the current conflict set, so that all known conflicting addresses are accessed first, ensuring that the next cache miss will occur on a different conflicting address. The process is repeated, issuing loads from all input memory addresses, incrementally finding conflicting addresses, one by one. Memory addresses that conflict in the cache belong to the same partition, whereas memory addresses belonging to different partitions do not conflict.

Claims

1. A method for managing memory in a computer comprising, for each of at least one input set of memory addresses, and iteratively for each of the memory addresses in the at least one input set:

a) loading a current memory address, belonging to the at least one input set, into a processor cache;

b) detecting whether an eviction occurs from the cache as a result of the loading of current memory address into the processor cache;

c) if an eviction is detected;

i) adding the current memory address to a conflict set of memory addresses;

ii) flushing the cache; and

iii) loading the conflict set into the cache;

d) if no eviction is detected, determining whether all of the input set of memory addresses has been loaded into the cache and, if not, loading a next one of the input set of memory addresses into a processor cache;

whereby the input set of memory addresses is determined to be partitioned into a plurality of partitions such that memory addresses that conflict in the cache belong to the same partition, whereas memory addresses belonging to different partitions do not conflict in the cache.

2. The method of claim 1, further comprising repeating the operations a)-d) of claim 1 for each of a plurality of input sets.

3. The method of claim 2, further comprising:

marking memory addresses in the current set as being used after they have been loaded; and

loading as current memory addresses of subsequent input sets only memory addresses not marked as used.

4. The method of claim 1, further comprising detecting whether an eviction occurs by detecting a change in a hardware performance counter.

5. The method of claim 1, further comprising:

configuring a hardware performance counter to count cache misses;

detecting whether an eviction occurs as a result of accessing a newly loaded memory address by reloading into the cache, one by one and in order, previously loaded memory addresses, examining the hardware performance counter after each reloading, and determining the eviction as being the result of loading the newly loaded memory address if the hardware performance counter indicates a cache miss upon loading of the newly loaded memory address.

6. The method of claim 1, further comprising detecting whether an eviction occurs as a function of access latency for each respective current memory address.

7. The method as in claim 1, in which the partitions have a cache-line granularity.

8. A system for managing memory in a computer comprising:

a processor;

a processor cache;

a cache partitioning module being configured, for each of at least one input set of memory addresses, and iteratively for each of the memory addresses in the at least one input set, for;

a) loading a current memory address, belonging to the at least one input set, into a processor cache;

b) detecting whether an eviction occurs from the cache as a result of the loading of current memory address into the processor cache;

c) if an eviction is detected;

i) adding the current memory address to a conflict set of memory addresses;

ii) flushing the cache; and

iii) loading the conflict set into the cache;

d) if no eviction is detected, determining whether all of the input set of memory addresses has been loaded into the cache and, if not, loading a next one of the input set of memory addresses into a processor cache;

whereby the input set of memory addresses is determined to be partitioned into a plurality of partitions such that memory addresses that conflict in the cache belong to the same partition, whereas memory addresses belonging to different partitions do not conflict in the cache.

9. The system of claim 8, in which the cache partitioning module is configured to repeat the operations a)-d) of claim 8 for each of a plurality of input sets.

10. The system of claim 9, in which the cache partitioning module is further configured:

for marking memory addresses in the current set as being used after they have been loaded; and

for loading as current memory addresses of subsequent input sets only memory addresses not marked as used.

11. The system as in claim 8, further comprising a hardware performance counter, said cache partitioning module being configured for detecting whether an eviction occurs by detecting a change in the hardware performance counter.

12. The system of claim 8, further comprising:

a hardware cache miss counter;

said cache partitioning module being configured for detecting whether an eviction occurs as a result of accessing a newly loaded memory address by reloading into the cache, one by one and in order, previously loaded memory addresses, examining the hardware cache miss counter after each reloading, and determining the eviction as being the result of loading the newly loaded memory address if the hardware cache miss counter indicates a cache miss upon loading of the newly loaded memory address.

13. The system of claim 8, in which the cache partitioning module is configured for detecting whether an eviction occurs as a function of access latency for each respective current memory address.

14. A non-transitory computer-readable storage medium storing instructions, the instructions, when executed by a processor, causing the processor:

for each of at least one input set of memory addresses, and iteratively for each of the memory addresses in the at least one input set;

a) to load a current memory address, belonging to the at least one input set, into a processor cache;

b) to detect whether an eviction occurs from the cache as a result of the loading of current memory address into the processor cache;

c) if an eviction is detected;

i) to add the current memory address to a conflict set of memory addresses;

ii) to flush the cache; and

iii) to load the conflict set into the cache; and

d) if no eviction is detected, to determine whether all of the input set of memory addresses has been loaded into the cache and, if not, to load a next one of the input set of memory addresses into a processor cache;

whereby the input set of memory addresses is determined to be partitioned into a plurality of partitions such that memory addresses that conflict in the cache belong to the same partition, whereas memory addresses belonging to different partitions do not conflict in the cache.

15. The medium of claim 14, said instructions, upon execution by the processor, further causing the processor to repeat the operations a)-d) of claim 14 for each of a plurality of input sets.

16. The medium of claim 15, said instructions, upon execution by the processor, further causing the processor:

to mark memory addresses in the current set as being used after they have been loaded; and

to load as current memory addresses of subsequent input sets only memory addresses not marked as used.

17. The medium of claim 14, said instructions, upon execution by the processor, further causing the processor to detect whether an eviction occurs by detecting a change in a hardware performance counter.

18. The medium of claim 14, in which the instructions, upon execution by the processor, further cause the processor:

to configure a hardware performance counter to count cache misses; and

to detect whether an eviction occurs as a result of accessing a newly loaded memory address by reloading into the cache, one by one and in order, previously loaded memory addresses, examining the hardware performance counter after each reloading, and determining the eviction as being the result of loading the newly loaded memory address if the hardware performance counter indicates a cache miss upon loading of the newly loaded memory address.

19. The medium of claim 14, in which the instructions, upon execution by the processor, further cause the processor to detect whether an eviction occurs as a function of access latency for each respective current memory address.

20. The medium of claim 14, in which the partitions have a cache-line granularity.

21. A computer program product having a non-transitory computer-readable storage medium storing computer-executable code, the code comprising a cache partitioning module configured, for each of at least one input set of memory addresses, and iteratively for each of the memory addresses in the at least one input set:

a) to load a current memory address, belonging to the at least one input set, into a processor cache;

b) to detect whether an eviction occurs from the cache as a result of the loading of current memory address into the processor cache;

c) if an eviction is detected;

i) to add the current memory address to a conflict set of memory addresses;

ii) to flush the cache; and

iii) to load the conflict set into the cache; and

d) if no eviction is detected, to determine whether all of the input set of memory addresses has been loaded into the cache and, if not, to load a next one of the input set of memory addresses into a processor cache;

whereby the input set of memory addresses is determined to be partitioned into a plurality of partitions such that memory addresses that conflict in the cache belong to the same partition, whereas memory addresses belonging to different partitions do not conflict in the cache.

22. The computer program product of claim 21, in which the cache partitioning module is further configured to repeat the operations a)-d) of claim 21 for each of a plurality of input sets.

23. The computer program product of claim 22, in which the cache partitioning module is further configured:

to mark memory addresses in the current set as being used after they have been loaded; and

to load as current memory addresses of subsequent input sets only memory addresses not marked as used.

24. The computer program product of claim 21, in which the cache partitioning module is further configured to detect whether an eviction occurs by detecting a change in a hardware performance counter.

25. The computer program product of claim 21, in which the cache partitioning module is further configured to:

to configure a hardware performance counter to count cache misses; and

to detect whether an eviction occurs as a result of accessing a newly loaded memory address by reloading into the cache, one by one and in order, previously loaded memory addresses, examining the hardware performance counter after each reloading, and determining the eviction as being the result of loading the newly loaded memory address if the hardware performance counter indicates a cache miss upon loading of the newly loaded memory address.

26. The computer program product of claim 21, in which the cache partitioning module is further configured to detect whether an eviction occurs as a function of access latency for each respective current memory address.

27. The computer program product of claim 21, in which the partitions have a cache-line granularity.