1. Improving Energy Efficiency by Making DRAM Less Randomly Accessed Hai Huang, Kang G. Shin, Charles Lefurgy, Tom Keller University of Michigan IBM Austin Research Lab

2. Overview Continual increase in the power budget allocated to main memory (i.e., DRAM)  E.g., in a mid-range IBM eServer system, 40% of the total system energy is consumed by its main memory subsystem By passively monitoring memory traffic and managing the power, existing power management techniques are not fully exploiting deeper power-saving states => Actively shape memory traffic to enable existing techniques to save more energy

3. Passive Monitoring Memory Traffic Why is passively monitoring memory traffic inefficient?  Memory accesses are random – good for performance, bad for energy consumption!  Idle time between consecutive memory accesses is often too short for use of the deeper power-saving state  Randomness is mostly due to OS’s arbitrary virtual-to-physical mapping

4. Example: Active vs. Passive Rank 0 Rank 1 Rank 0 Rank 1 time Active memory traffic management High-powerLow-powerUltra Low-power time Passive memory traffic management

5. How to Shape Memory Traffic Essentially, we need to artificially create disparity in access frequency among different memory ranks Hot Ranks and Cold Ranks Disparity in access frequency can be created by finding and migrating frequently-accessed pages to a subset of memory ranks  Hot ranks: contain frequently-accessed pages  Cold ranks: contain infrequently-accessed and unmapped pages Page migration can be done by system software

6. Implementation page counter MC Rank 0 Rank 1 Rank 2 Rank 3 Hot ranks Cold ranks Operating System Migration thread Time triggers Migrate (old_page, new_page) Second level page table Process First level page table Modify PT

7. Issues with Page Migration There is a cost associated with each page migration Memory access frequency Is often highly skewed!!! 6% pages causes 75% accesses 14% pages causes 90% accesses Not all pages need to be migrated

8. Evaluation Simulators  Mambo [IBM] – A full-machine simulator, cycle-accurate, supports PowerPC architecture  Memsim [IBM] – Detailed trace-driven main memory simulator, written in CSIM Workloads  Low memory-intensive workload: SPECjbb + bzip + crafty  High memory-intensive workload: SPECjbb + art + mcf SPECjbb: simulating 8 warehouses SPEC2000 benchmarks: using Reference input set

9. Low Memory-Intensive Workload

10. High Memory-Intensive Workload

11. Summary of Results Energy:  Actively shaping memory traffic saves 35% more energy than passively monitoring Performance :  Low memory-intensive workload: small impact on performance  High memory-intensive workload: significantly degrades performance due to more contention on hot ranks Cost :  Use hardware counters, or  Software page faults

12. Conclusion Actively shaping memory traffic allows existing power management techniques to more effectively save power Highly-skewed page accesses are observed Alternative main memory design:  Use high-performance/highly-parallel ranks as hot ranks  Use low-performance/low-power ranks as cold ranks Allows frequently-accessed pages to be accessed faster Allows memory ranks that hold infrequently-accessed and unmapped pages to consume less energy