Gaining Insights into Multi-Core Cache Partitioning: Bridging the Gap between Simulation and Real Systems Jiang Lin1, Qingda Lu2, Xiaoning Ding2, Zhao Zhang1, Xiaodong Zhang2, and P. Sadayappan2 1 Department of ECE Iowa State University 2 Department of CSE The Ohio State University Shared Caches Can be a Critical Bottleneck in Multi-Core Processors L2/L3 caches are shared by multiple cores Intel Xeon 51xx (2core/L2) AMD Barcelona (4core/L3) Sun T2, ... (8core/L2) Core Core …… Core Shared L2/L3 cache Effective cache partitioning is critical to address the bottleneck caused by the conflicting accesses in shared caches. Several hardware cache partitioning methods have been proposed with different optimization objectives Performance: [HPCA’02], [HPCA’04], [Micro’06] Fairness: [PACT’04], [ICS’07], [SIGMETRICS’07] QoS: [ICS’04], [ISCA’07] 2 Limitations of Simulation-Based Studies Excessive simulation time Whole programs can not be evaluated. It would take several weeks/months to complete a single SPEC CPU2006 benchmark As the number of cores continues to increase, simulation ability becomes even more limited Absence of long-term OS activities Interactions between processor/OS affect performance significantly Proneness to simulation inaccuracy Bugs in simulator Impossible to model many dynamics and details of the system 3 Our Approach to Address the Issues Design and implement OS-based Cache Partitioning Embedding cache partitioning mechanism in OS By enhancing page coloring technique To support both static and dynamic cache partitioning Evaluate cache partitioning policies on commodity processors Execution- and measurement-based Run applications to completion Measure performance with hardware counters 4 Four Questions to Answer Can we confirm the conclusions made by the simulation- based studies? Can we provide new insights and findings that simulation is not able to? Can we make a case for our OS-based approach as an effective option to evaluate multicore cache partitioning designs? What are advantages and disadvantages for OS-based cache partitioning? 5 Outline Introduction Design and implementation of OS-based cache partitioning mechanisms Evaluation environment and workload construction Cache partitioning policies and their results Conclusion 6 OS-Based Cache Partitioning Mechanisms Static cache partitioning Predetermines the amount of cache blocks allocated to each program at the beginning of its execution Page coloring enhancement Divides shared cache to multiple regions and partition cache regions through OS page address mapping Dynamic cache partitioning Adjusts cache quota among processes dynamically Page re-coloring Dynamically changes processes’ cache usage through OS page address re-mapping 7 Page Coloring •Physically indexed caches are divided into multiple regions (colors). •All cache lines in a physical page are cached in one of those regions (colors). Physically indexed cache Virtual address virtual page number OS control Address translation Physical address physical page number page offset …… Page offset = OS can control the page color of a virtual page through address mapping (by selecting a physical page with a specific value in its page color bits). Cache address 8 Cache tag Set index page color bits Block offset Enhancement for Static Cache Partitioning Physical pages are grouped to page bins according to their page color 1 2 3 4 … …… … … OS address mapping i i+1 i+2 Physically indexed cache … Shared cache is partitioned between two…… processes through address mapping. Process 1 ... 1 2 3 4 …… Cost: Main memory space needs to be partitioned too (co-partitioning). … i i+1 i+2 … 9 Process 2 … …… … …… Dynamic Cache Partitioning Why? Programs have dynamic behaviors Most proposed schemes are dynamic How? Page re-coloring How to handle overhead? Measure overhead by performance counter Remove overhead in result (emulating hardware schemes) 10 Dynamic Cache Partitioning through Page Re-Coloring Allocated Allocated color color 0 1 2 Pages of a process are organized into linked lists by their colors. Memory allocation guarantees that pages are evenly distributed into all the lists (colors) to avoid hot points. 3 …… N-1 11 page links table Page re-coloring: Allocate page in new color Copy memory contents Free old page Control the Page Migration Overhead Control the frequency of page migration Frequent enough to capture application phase changes Not too often to introduce large page migration overhead Lazy migration: avoid unnecessary page migration Observation: Not all pages are accessed between their two migrations. Optimization: do not migrate a page until it is accessed 12 Lazy Page Migration Allocated Allocated color color 0 1 2 3 …… After the optimization N-1 13 Avoid unnecessary page migration for these pages! Process page links On average, 2% page migration overhead Up to 7%. Outline Introduction Design and implementation of OS-based cache partitioning mechanisms Evaluation environment and workload construction Cache partitioning policies and their results Conclusion 14 Experimental Environment Dell PowerEdge1950 Two-way SMP, Intel dual-core Xeon 5160 Shared 4MB L2 cache, 16-way 8GB Fully Buffered DIMM Red Hat Enterprise Linux 4.0 2.6.20.3 kernel Performance counter tools from HP (Pfmon) Divide L2 cache into 16 colors 15 Benchmark Classification 6 9 6 8 29 benchmarks from SPEC CPU2006 Is it sensitive to L2 cache capacity? Red group: IPC(1M L2 cache)/IPC(4M L2 cache) < 80% Give red benchmarks more cache: big performance gain Yellow group: 80% <IPC(1M L2 cache)/IPC(4M L2 cache) < 95% Give yellow benchmarks more cache: moderate performance gain Else: Does it extensively access L2 cache? Green group: > = 14 accesses / 1K cycle Give it small cache Black group: < 14 accesses / 1K cycle 16 Cache insensitive Workload Construction 2-core 6 6 RR (3 pairs) 9 RY (6 pairs) YY (3 pairs) 6 RG (6 pairs) YG (6 pairs) 9 6 GG (3 pairs) 27 workloads: representative benchmark combinations 17 Outline Introduction OS-based cache partitioning mechanism Evaluation environment and workload construction Cache partitioning policies and their results Performance Fairness Conclusion 18 Performance – Metrics Divide metrics into evaluation metrics and policy metrics [PACT’06] Evaluation metrics: Optimization objectives, not always available during run-time Policy metrics Used to drive dynamic partitioning policies: available during run-time Sum of IPC, Combined cache miss rate, Combined cache misses 19 Static Partitioning Total #color of cache: 16 Give at least two colors to each program Make sure that each program get 1GB memory to avoid swapping (because of co-partitioning) Try all possible partitionings for all workloads (2:14), (3:13), (4:12) ……. (8,8), ……, (13:3), (14:2) Get value of evaluation metrics Compared with performance of all partitionings with performance of shared cache 20 Performance – Optimal Static Partitioning Performance gain of optimal static partitioning Throughtput Average Weighted Speedup Normalized SMT Speedup Fair Speedup 1.25 1.20 1.15 1.10 1.05 1.00 RR RY RG YY YG GG Confirm that cache partitioning has significant performance impact Different evaluation metrics have different performance gains RG-type of workloads have largest performance gains (up to 47%) Other types of workloads also have performance gains (2% to 10%) 21 A New Finding Workload RG1: 401.bzip2 (Red) + 410.bwaves (Green) Intuitively, giving more cache space to 401.bzip2 (Red) Increases the performance of 401.bzip2 largely (Red) Decreases the performance of 410.bwaves slightly (Green) However, we observe that 22 Insight into Our Finding 23 Partitionings Partitionings 13:3 13:3 14:2 14:2 11:5 11:5 12:4 12:4 9:7 9:7 10:6 10:6 7:9 7:9 8:8 8:8 6:10 6:10 5:11 5:11 4:12 4:12 3.05 156 154 3.00 152 2.95 150 2.90 148 2.85 146 2.80 144 2.75 142 2.70 140 2:14 2:14 3:13 3:13 GB/s ns Memory Utilization AverageBandwidth Memory Access Latency Insight into Our Finding We have the same observation in RG4, RG5 and YG5 This is not observed by simulation Did not model main memory sub-system in detail Assumed fixed memory access latency Shows the advantages of our execution- and measurement- base study 24 Performance - Dynamic Partition Policy Init: Partition the cache as (8:8) finished No Yes Exit Run current partition (P0:P1) for one epoch Try one epoch for each of the two neighboring partitions: (P0 – 1: P1+1) and (P0 + 1: P1-1) A simple greedy policy. Emulate policy of [HPCA’02] Choose next partitioning with best policy metrics measurement 25 Performance – Static & Dynamic Use combined miss rates as policy metrics For RG-type, and some RY-type: Static partitioning outperforms dynamic partitioning For RR- and RY-type, and some RY-type Dynamic partitioning outperforms static partitioning 26 Fairness – Metrics and Policy [PACT’04] Metrics Evaluation metrics FM0 difference in slowdown, small is better Policy metrics Policy Repartitioning and rollback 27 Fairness - Result Dynamic partitioning can achieve better fairness If we use FM0 as both evaluation metrics and policy metrics None of policy metrics (FM1 to FM5) is good enough to drive the partitioning policy to get comparable fairness with static partitioning Strong correlation was reported in simulation-based study – [PACT’04] None of policy metrics has consistently strong correlation with FM0 SPEC CPU2006 (ref input) SPEC CPU2000 (test input) Complete trillions of instructions less than one billion instruction 4MB L2 cache 512KB L2 cache 28 Conclusion Confirmed some conclusions made by simulations Provided new insights and findings Give cache space from one to another, increase performance of both Poor correlation between evaluation and policy metrics for fairness Made a case for our OS-based approach as an effective option for evaluation of multicore cache partitioning Advantages of OS-based cache partitioning Working on commodity processors for an execution- and measurement-based study Disadvantages of OS-based cache partitioning Co-partitioning (may underutilize memory), migration overhead 29 Ongoing Work Reduce migration overhead on commodity processors Cache partitioning at the compiler level Partition cache at object level Hybrid cache partitioning method Remove the cost of co-partitioning Avoid page migration overhead 30 Thanks! Gaining Insights into Multi-Core Cache Partitioning: Bridging the Gap between Simulation and Real Systems Jiang Lin1, Qingda Lu2, Xiaoning Ding2, Zhao Zhang1, Xiaodong Zhang2, and P. Sadayappan2 1 Iowa State University 2 The Ohio State University Backup Slides 32 Fairness - Correlation between Evaluation Metrics and Policy Metrics (Reported by [PACT’04]) 1 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 -1 apsi+equake gzip+apsi Corr(M1,M0) Corr(M4,M0) swim+gzip Corr(M2,M0) Corr(M5,M0) tree+mcf AVG18 Corr(M3,M0) Strong correlation was reported in simulation study – [PACT’04] 33 Fairness - Correlation between Evaluation Metrics and Policy Metrics (Our result) 1 0.8 FM1 0.6 0.4 0.2 FM3 0 -0.2 -0.4 -0.6 FM5 FM4 -0.8 -1 YY1 YY2 YY3 YG1 YG2 YG3 YG4 YG5 YG6 GG1 GG2 None of policy metrics has consistently strong correlation with FM0 SPEC CPU2006 (ref input) SPEC CPU2000 (test input) Complete trillions of instructions less than one billion instruction 4MB L2 cache 512KB L2 cache 34 GG3