Eta Scale Ab

Samsung Electronics Co., Ltd.

Other

Hitachi, Ltd.

Yissum Research Development Company of the Hebrew University of Jerusalem Ltd.

Universitéde Liége

Stefanos Kaxiras

Alberto Ros

Erik Hagersten

David Black-Schaffer

Andreas Sembrant

Christos Sakalis

Alexandra Jimborean

Magnus Själander

Erik Ernst Hagersten

Françoise Remacle

Raphael D Levine

A computer system for dynamic enforcement of store atomicity includes multiple processor cores, local cache memory for each processor core, a shared memory, a separate store buffer for each processor core for executed stores that are not yet performed and a coherence mechanism. A first processor core load on a first processor core receives a value at a first time from a first processor core store in the store buffer and prevents any other first processor core load younger than the first processor core load in program order from committing until a second time when the first processor core store is performed. Between the first time and the second time any load younger in program load than the first processor core load and having an address matched by coherence invalidation or an address matched by an eviction is squashed.

System and method for dynamic enforcement of store atomicity

Synchronization events associated with cache coherence are monitored without using invalidations. A callback-read is issued to a memory address associated with the synchronization event, which callback-read either reads the last value written in the memory address or blocks until a next write takes place in the memory address and reads a newly written value.

System and method for event monitoring in cache coherence protocols without explicit invalidations

Methods and systems for self-invalidating cachelines in a computer system having a plurality of cores are described. A first one of the plurality of cores, requests to load a memory block from a cache memory local to the first one of the plurality of cores, which request results in a cache miss. This results in checking a read-after-write detection structure to determine if a race condition exists for the memory block. If a race condition exists for the memory block, program order is enforced by the first one of the plurality of cores at least between any older loads and any younger loads with respect to the load that detects the prior store in the first one of the plurality of cores that issued the load of the memory block and causing one or more cache lines in the local cache memory to be self-invalidated.

System and method for self-invalidation, self-downgrade cachecoherence protocols

A system and method for efficiently preventing visible side-effects in the memory hierarchy during speculative execution is disclosed. Hiding the side-effects of executed instructions in the whole memory hierarchy is both expensive, in terms of performance and energy, and complicated. A system and method is disclosed to hide the side-effects of speculative loads in the cache(s) until the earliest time these speculative loads become non-speculative. A refinement is disclosed where loads that hit in the L1 cache are allowed to proceed by keeping their side-effects on the L1 cache hidden until these loads become non-speculative, and all other speculative loads that miss in the cache(s) are prevented from executing until they become non-speculative. To limit the performance deterioration caused by these delayed loads, a system and method is disclosed that augments the cache(s) with a value predictor or a re-computation engine that supplies predicted or recomputed values to the loads that missed in the cache(s).

Systems and methods for invisible speculative execution

A method for performing store buffer coalescing in a multiprocessor computer system includes forming, in a coalescing store buffer associated with a core in said multiprocessor system, an atomic group of writes; and performing each individual write in said atomic group in an order which is a function of an address in a memory system to which each of the writes in said atomic group are being written.

Systems and methods for non-speculative store coalescing and generating atomic write sets using address subsets

Methods and systems for maintaining validity of a memory model in a multiple core computer system are described. A first core prevents a store instruction from being performed by another core until a condition is met which enables reordered instructions to validly execute.

System and method for non-speculative reordering of load accesses

According to embodiments described herein, the hierarchical complexity for coherence protocols associated with clustered cache architectures can be encapsulated in a simple function, i.e., that of determining when a data block is shared entirely within a cluster (i.e., a sub-tree of the hierarchy) and is private from the outside. This allows embodiments to eliminate complex recursive coherence operations that span the hierarchy and instead employ simple coherence mechanisms such as self-invalidation and write-through but which are restricted to operate where a data block is shared. Thus embodiments recognize that, in the context of clustered cache hierarchies, data can be shared entirely within one cluster but can be private (unshared) to this cluster when viewed from the perspective of other clusters. This characteristic of the data can be determined and then used to locally simplify coherence protocols.

Multi-core computer systems with private/shared cache line indicators

Caches may be vulnerable to side-channel attacks, such as Spectre and Meltdown, that involve speculative execution of instructions, revealing information about a cache that the attacker is not permitted to access. Access permission may be stored in the cache, such as in an entry of a cache table or in the region information for a cache table. Optionally, the access permission may be re-checked if the access permission changes while a memory instruction is pending. Optionally, a random index value may be stored in a cache and used, at least in part, to identify a memory location of a cacheline. Optionally, cachelines that are involved in speculative loads for memory instructions may be marked as speculative. On condition of resolving the speculative load as non-speculative, the cacheline may be marked as non-speculative; and on condition of resolving the speculative load as mis-speculated, the cacheline may be removed from the cache.

System protecting caches from side-channel attacks

Methods and systems which, for example, reduce energy usage in cache memories are described. Cache location information regarding the location of cachelines which are stored in a tracked portion of a memory hierarchy is stored in a cache location table. Address tags are stored with corresponding location information in the cache location table to associate the address tag with the cacheline and its cache location information. When a cacheline is moved to a new location in the memory hierarchy, the cache location table is updated so that the cache location information indicates where the cacheline is located within the memory hierarchy.

Systems and methods for reducing first level cache energy by eliminating cache address tags

Methods and systems for in direct data access in, e.g., multi-level cache memory systems are described. A cache memory system includes a cache location buffer configured to store cache location entries, wherein each cache location entry includes an address tag and a cache location table which are associated with a respective cacheline stored in a cache memory. The system also includes a first cache memory configured to store cachelines, each cacheline having data and an identity of a corresponding cache location entry in the cache location buffer, and a second cache memory configured to store cachelines, each cacheline having data and an identity of a corresponding cache location entry in the cache location buffer. Responsive to a memory access request for a cacheline, the cache location buffer generates access information using one of the cache location tables which enables access to the cacheline without performing a tag comparison at the one of the first and second cache memories.

Systems and methods for direct data access in multi-level cache memory hierarchies

A computer processing system includes a plurality of nodes, each node having at least one processor core and at least one level of cache memory which is private to the node, a shared, last level cache (LLC) memory device and a shared, last level cache location buffer containing cache location entries, each cache location entry storing an address tag and a plurality of location information. The location information stored in a cache location entry points to an identified cacheline location within the LLC that stores a cacheline associated with the location information. The cacheline stored in the LLC has associated information identifying the cache location entry.

Systems and methods for implementing a tag-less shared cache and a larger backing cache

Systems and methods for coherence in clustered cache hierarchies

System and methods for cache coherence in a multi-core processing environment having a local/shared cache hierarchy. The system includes multiple processor cores, a main memory, and a local cache memory associated with each core for storing cache lines accessible only by the associated core. Cache lines are classified as either private or shared. A shared cache memory is coupled to the local cache memories and main memory for storing cache lines. The cores follow a write-back to the local memory for private cache lines, and a write-through to the shared memory for shared cache lines. Shared cache lines in local cache memory enter a transient dirty state when written by the core. Shared cache lines transition from a transient dirty to a valid state with a self-initiated write-through to the shared memory. The write-through to shared memory can include only data that was modified in the transient dirty state.

Growing community of inventors

Uppsala, Sweden

Stefanos Kaxiras