Excited
Title: Some Considerations on Learning to Explore via Meta-Reinforcement Learning
Idea: Modify the mathmatical target so that take exploration into consideration, come up with 2 meta RL algorithms E-MAML and E-RL, and propose a Krazy World environment to benchmark meta RL.
Title: Using Classical Planners for Tasks with Continuous Operators in Robotics
Idea:
Contribution:
Experiment:
Title: Learning Generalized Reactive Policies using Deep Neural Networks
Idea:
Contribution:
Experiments:
Title: Transparent Offloading and Mapping (TOM):Enabling Programmer-Transparent Near-Data Processing in GPU Systems
Topic: Data offloading for multiple 3D memory-stacked GPGPU architechture to alleviate the memory bandwidth bottleneck.
Idea: alleviate the memory bandwitdh bottleneck mainly by 2 phases: 1) use compiler-based technique to determine potentially offload candidate by calculate the benifit from offload, and make the offload decision at runtime, 2) use a data mapping technique that doesn’t need programmers’ effort by combining both hardware/software then determine the mapping method based on a small period observation using memory access patterns.
Contribution: 1)propose a new data offloading mechanism that statically identifies instruction blocks benefit from offloading, and dynamically decides whether candidate can be offload in runtime 2) propose a new programmer-transparent data mapping mechanism exploits the predictability in memory access patterns of to co-locate offloaded code and data in the same memory stack 3)comprehensively evaluate the mechanism using 10 memory-intensive GPGPU applications across different system configurations
The result shows that their mechanism is practical and effective approach to enabling programmer-transparent near-data processing in GPU systems.
Title: GPU Performance and Power Estimation Using Machine Learning
Topic: Using machine learning for GPU performance and power rapid estimation
Idea: For training: Vary GPU setting from 3 perspective(They use 3 hardware params to illustrate the method, which can also be extend to many params): #compute units(CU), Mem Freq, Engine Freq to collect performance counters data, use K-means to cluster kernel performance: 1. CU scale, 2. Mem Freq & Engine Freq scale, use neural networks to classify kernel to cluster.
or prediction: run kernel on one setting once then classify to pattern cluster, use corresponding cluster centroid scaling values to make predictions on various GPU hardware setting.
Contribution: demonstrate on real hardware that the performance and power of GPGPU kernels scale in a limited number of ways as hardware configuration parameters are changed(finite unique scaling patterns), perform machine learning to estimate new kernel, describe an estimation model that run a kernel once on a hardware configuration then get performance and power over various configurations
The result shows that their accuracy is comparable to the accuracy of cycle-level simulators, and on real hardware the trained model runs equal or faster than native programs.
Title: Architectural Support for Address Translation on GPUs
Topic: explore GPU Memory Management Units (MMUs) consisting of Translation Lookaside Buffers (TLBs) and page table walkers (PTWs), analyze current technology and propose augments
Idea: In Heterogeneous Systems Architecture, the uniform virtual address of GPU/CPU memory can provide programming benefits and need efficient hardware, GPU’s warp-based execution model plays an important role. Using CPU-like MMU design would hurt GPU performance, they explored cache-conscious warp/wavefront scheduling for MMU design and how TLB affects dynamic warp formation. They explored and propose methods in 3 aspects: Address Translation for GPU, Cache Conscious Warp Scheduling, thread block compaction.
The benefit of their work: remove need of CPU to handle GPU MMU, support multi-context and support application libraries.
General sight: Default warp scheduling would break temporal locality, sophisticated warp scheduling would loss effectiveness.
Title: Vulnerable GPU Memory Management: Towards Recovering Raw Data from GPU
Topic: The security issue of Heterogeneous architecture from GPU memory management and algorighm to recover data from residue.
Idea: Analyze the weakness of HSA data security from GPU perspective, provide an algorighm that could recover image from GPU residue across the multiuser OS and Cloud computing, in 3 steps: get tilt, utilize the local similarity to idetify the size of image and remove leading blocks.
Title: Page Placement Strategies for GPUs within Heterogeneous Memory Systems
Topic: improve page placement performance for GPU in Heterogeneous Memory system
Idea: proposed bandwidth-aware (BW-AWARE) placement, maximizes GPU throughput by controlling the weights of BO and CO in aggregated memory, and optimize selectively place hot pages in the BO memory when capacity is limited, use profile to augment that provide information about hotness and size of data structures.
Contribution: a compiler-based GPU page placement solution
Title: APRES: Improving Cache Efficiency by Exploiting Load Characteristics on GPUs
Topic: Analyze the situation of data cache and improve GPU L1 cache performance based on group warp priority with prefetch to enhance cache reuse
Idea: Based on the locality of data usage, Locality Aware Warp Scheduler (LAWS) group warps who share last load PC and set priority to do greedy manage, store all global memory load on Last Load Table, put the grouping information in Warp Group Table and schedule execute order by the Load Store Unit (LSU) and Scheduling Aware Prefetching (SAP), who would put the missing warp at low priority and ask LAWS to give high priority to the prefetch warp for that missing group, with stored previous load and warp ID in the missing group.
Contribution: Analysis of load instructions on GPU and a mechanism named APRES that combine grouping warp to execute with priority and prefetching.
Intellectual merit: Figure out the locality of cache utilization and opportunity to prefetch, proposed dedicated mechanism aimed at GPU L1 cache optimization based on the observation: LAWS group warps to do priority management, SAP change missing warp to the end and do prefetch.
Strengths: Detailed legends as examples to illustrate the mechanism
Weakness: Performance may limited because of simulation and no real GPU environment
Title: Access Pattern-Aware Cache Management for Improving Data Utilization in GPU
Topic: A mechanism to improve GPU performance based on data usage pattern
Idea: The design of GPU memory hierarchy not support well for warp, previous warp-level cache mechanisms are too coarse that streaming cannot benefit and some instructions with strong temporal locality are forced to bypass. Detect locality data type by monitoring per load exemplar warp and use cache tag array to track data sharing, then use different strategies for the according locality data types, make corresponding hardware design to support the mechanism. For streaming type bypass the L1 cache and directly use L2, for inter-warp utilize the LRU stategy, for intra-warp use a protection algorithm that estimate the lifespan of data to utilize the locality until mostly end the reuse.
Contribution: Analysis of the weakness of previous GPU cache management. Proposed a hardware based solution for improve GPU L1 cache performance.
Intellectual merit: Use the notion of access pattern similarity (APS) to measure the consistency of access patterns. Analysis the locality data type for per-load level optimization. Use a protection algorithm for utilize the dominant intra-warp locality.
Strengths: Clearly illustrate the strategies and design for APCM, detailed experiments to show the performance.
Weakness: Heuristic based that the locality data types tend to not change.