Cloud datacenters are underutilized due to server over-provisioning. To increase datacenter utilization, cloud providers offer users an option to run workloads such as big data analytics on the underutilized resources, in the form of cheap yet revocable transient servers (e.g., EC2 spot instances, GCE preemptible instances). Though at highly reduced prices, deploying big data analytics on the unstable cloud transient servers can severely degrade the job performance due to instance revocations. To tackle this issue, this paper proposes iSpot, a cost-effective transient server provisioning framework for achieving predictable performance in the cloud, by focusing on Spark as a representative Directed Acyclic Graph (DAG)-style big data analytics workload. It first identifies the stable cloud transient servers during the job execution by devising an accurate Long Short-Term Memory (LSTM)-based price prediction method. Leveraging automatic job profiling and the acquired DAG information of stages, we further build an analytical performance model and present a lightweight critical data checkpointing mechanism for Spark, to enable our design of iSpot provisioning strategy for guaranteeing the job performance on stable transient servers. Extensive prototype experiments on both EC2 spot instances and GCE preemptible instances demonstrate that, iSpot is able to guarantee the performance of big data analytics running on cloud transient servers while reducing the job budget by up to 83.8% in comparison to the state-of-the-art server provisioning strategies, yet with acceptable runtime overhead.

Processing-In-Memory (PIM) has emerged as a high-performance and energy-efficient computing paradigm for accelerating convolutional neural network (CNN) applications. Resistive random access memory (ReRAM) has been widely used in PIM architectures due to its extremely high efficiency for accelerating matrix-vector multiplications through analog computing. However, because CNN training usually requires high-precision computation in the backward propagation (BP) stage, the limited precision of analog PIM accelerators impedes their adoption in CNN training. In this paper, we propose ReHy, a hybrid PIM accelerator to support CNN training in ReRAM arrays. It is composed of Analog PIM (APIM) and Digital PIM (DPIM) modules. ReHy uses APIM to accelerate the feed-forward propagation (FP) stage for high performance, and DPIM to process the BP stage for high accuracy. We exploit the capability of ReRAM for Boolean logic operations to design the DPIM architecture. Particularly, we design floating-point multiplication and addition operators to support matrix multiplications in ReRAM arrays. We also propose a performance model to offload high-precision matrix multiplications to DPIM according to the data parallelism. Experimental results show that ReHy can speed up CNN training by 18.8 and 2.3, and reduce energy consumption by 21.2 and 3.2, compared with CPUs and state-of-the-art FloatPIM, respectively.

Hybrid memory systems composed of emerging non-volatile memory (NVM) and DRAM have drawn increasing attention in recent years. To fully exploit the advantages of both NVM and DRAM, a primary goal is to properly place application data on the hybrid memories. Previous studies have focused on page migration schemes to achieve higher performance and energy efficiency. However, those schemes all rely on online page access monitoring (costly), and data migration at the page granularity may cause additional overhead due to DRAM bandwidth contention and maintenance of cache/TLB consistency. In this paper, we present Object-level memory Allocation and Migration (OAM) mechanisms for hybrid memory systems. OAM exploits a profiling tool to characterize objects' memory access patterns at different execution phases of applications, and applies a performance/energy model to direct the initial static memory allocation and runtime dynamic object migration between NVM and DRAM. Based on our newly-developed programming interfaces for hybrid memory systems, application source codes can be automatically transformed via static code instrumentation. We evaluate OAM on an emulated hybrid memory system, and experimental results show that OAM can significantly reduce system energy-delay-product by 61% on average compared to a page-interleaving data placement scheme. It can also significantly reduce data migration overhead by 83% and 69% compared to the state-of-the-art page migration scheme CLOCK-DWF and 2PP, respectively, while improving application performance by up to 22% and 10%.

Many iterative graph processing systems have recently been developed to analyze graphs. Although they are effective from different aspects, there is an important issue that has not been addressed yet. A real-world graph follows the power-law property, in which a small number of vertices have high degrees (i.e., are connected to most other vertices in the graph). These vertices are called hot-vertices and usually require more iterations to converge. In the existing solutions, these hot-vertices may be allocated to many or even all graph partitions along with other vertices that are easy to converge. As the result, the partitions with hot-vertices have to be loaded repeatedly (and consequently the system suffers from high data access cost), although perhaps only a few vertices in these partitions are active. To cope with this issue, we develop an efficient open source graph partition manager, called GGraph, which can be integrated into the existing graph processing systems to efficiently support iterative graph processing, by taking into account the power-law property of the graph structure. It uses a novel graph repartitioning scheme with low overhead to dynamically partition the hot-vertices together, so as to avoid loading the inactive vertices in the same partition as the repeatedly processed hot-vertices. By such means, it not only enables less data access cost, but also enables the privileged processing of the hot-vertices. In order to further increase the convergence speed, a scheduling algorithm is further proposed in this work to prioritize the processing of the hot-vertices with low overhead. To demonstrate the efficiency of GGraph, we plug it into four state-of-the-art graph processing systems, i.e., Gemini, GraphChi, Chaos, and GridGraph, and experimental results show that GGraph improves their performance by up to 3.2 times, 3.8 times, 3.9 times, 3.5 times, respectively.