Scalable hardware memory disambiguation

This dissertation deals with one of the long-standing problems in Computer Architecture – the problem of memory disambiguation. Microprocessors typically reorder memory instructions during execution to improve concurrency. Such microprocessors use hardware memory structures for memory disambiguation, known as LoadStore Queues (LSQs), to ensure that memory instruction dependences are satisfied even when the memory instructions execute out-of-order. A typical LSQ implementation (circa 2006) holds all in-flight memory instructions in a physically centralized LSQ and performs a fully associative search on all buffered instructions to ensure that memory dependences are satisfied. These LSQ implementations do not scale because they use large, fully associative structures, which are known to be slow and power hungry. The increasing trend towards distributed microarchitectures further exacerbates these problems. As on-chip wire delays increase and high-performance processors become necessarily distributed, centralized structures such as the LSQ can limit scalability. This dissertation describes techniques to create scalable LSQs in both centralized and distributed microarchitectures. The problems and solutions described in this thesis are motivated and validated by real system designs. The dissertation starts with a description of the partitioned primary memory system of the TRIPS processor, of which the LSQ is an important component, and then through a series of optimizations describes how the power, area, and centralization problems of the LSQ can be solved with minor performance losses (if at all) even for large number of in flight memory instructions. The four solutions described in this dissertation — partitioning, filtering, late binding and efficient overflow management — enable power-, area-efficient, distributed and scalable LSQs, which in turn enable aggressive large-window processors capable of simultaneously executing thousands of instructions. To mitigate the power problem, we replaced the power-hungry, fully associative search with a power-efficient hash table lookup using a simple address-based Bloom filter. Bloom filters are probabilistic data structures used for testing set membership and can be used to quickly check if an instruction with the same data address is likely to be found in the LSQ without performing the associative search. Bloom filters typically eliminate more than 80% of the associative searches and they are highly effective because in most programs, it is uncommon for loads and stores to have the same data address and be in execution simultaneously. To rectify the area problem, we observe the fact that only a small fraction of all memory instructions are dependent, that only such dependent instructions need to be buffered in the LSQ, and that these instructions need to be in the LSQ only for certain parts of the pipelined execution. We propose two mechanisms to exploit these observations. The first mechanism, area filtering, is a hardware mechanism that couples Bloom filters and dependence predictors to dynamically identify and buffer only those instructions which are likely to be dependent. The second mechanism, late binding, reduces the occupancy and hence size of the LSQ. Both of these optimizations allows the number of LSQ slots to be reduced by up to one-half compared to a traditional organization without any performance degradation. Finally, we describe a new decentralized LSQ design for handling LSQ structural hazards in distributed microarchitectures. Decentralization of LSQs, and to a large extent distributed microarchitectures with memory speculation, has proved to be impractical because of the high performance penalties associated with the mechanisms for dealing with hazards. To solve this problem, we applied classic flow-control techniques from interconnection networks for handling resource con- flicts. The first method, memory-side buffering, buffers the overflowing instructions in a separate buffer near the LSQs. The second scheme, execution-side NACKing, sends the overflowing instruction back to the issue window from which it is later re-issued. The third scheme, network buffering, uses the buffers in the interconnection network between the execution units and memory to hold instructions when the LSQ is full, and uses virtual channel flow control to avoid deadlocks. The network buffering scheme is the most robust of all the overflow schemes and shows less than 1% performance degradation due to overflows for a subset of SPEC CPU 2000 and EEMBC benchmarks on a cycle-accurate simulator that closely models the TRIPS processor. The techniques proposed in this dissertation are independent, architectureneutral and their cumulative benefits result in LSQs that can be partitioned at a fine granularity and have low design complexity. Each of these partitions selectively buffers only memory instructions with true dependences and can be closely coupled with the execution units thus minimizing power, area, and latency. Such LSQ designs with near-ideal characteristics are well suited for microarchitectures with thousands of instructions in-flight and may enable even more aggressive microarchitectures in the future.