Chulmin Kim

OPAMP: Evaluation Framework for Optimal Page Allocation of Hybrid Main Memory Architecture

Main memory as a hybrid between DRAM and nonvolatile memory is rapidly considered as a basic building block of computing systems. Despite widely-performed researches no one can confirm whether hybrid memory is at its full performance in terms of energy consumption, time delay or both. The main problem is that evaluating their performance in comparison with the optimal performance is challenging since deriving the optimal value is NP-complete. In this paper, we design and implement an evaluation framework termed OPAMP, which calculates optimal performance of the hybrid memory environment. This system gathers workload, specification of DRAM and PRAM, and environmental parameters of the hybrid main memory. After that, it calculates the maximum performance under the corresponding conditions. We suggest the way of deriving the optimal value by profiling instead of page migration which is the mainstream of recent researches on hybrid main memory system. Also, proportion of DRAMs size to PRAMs and proportion of DRAMs usage space to PRAMs are impactive factors. While designing hybrid main memory, those two variables must be determined carefully and OPAMP gives the guideline to the researchers.

GHOST: GPGPU-offloaded high performance storage I/O deduplication for primary storage system

Data deduplication has been an effective way to eliminate redundant data mainly for backup storage systems. Since the recent primary storage systems in cloud services are expected to have the redundancy, the deduplication technique can also bring significant cost saving for the primary storage. However, the primary storage system requires high performance requirement about several GBs per second. Most conventional deduplication techniques targeted the performance requirement of 200-300MB/s.n In an attempt to achieve a high performance storage deduplication system at the primary storage, we thoroughly analyze the performance bottleneck of previous deduplication systems to enhance the system to meet the requirement of the primary storage. The new performance bottleneck of deduplication in the primary storage lies on not only key-value store lookup, also computation for data segmentation and fingerprinting due to recent technology improvement of flash devices such as SSD. To overcome the bottlenecks, we propose a new deduplication system utilizing GPGPU. Our proposed system, termed GHOST, includes the followings to offload and optimize the deduplication processing in GPGPU: (1) In-Host Data Cache, (2) Destage-aware Data offloading to GPGPU and (3) In-GPGPU Table Cache of key-value store. These techniques improve the offloaded deduplication processing about 10-20% on the reasonable workload of the primary storage compared to the naive approach. Our proposed deduplication system can achieve 1.5GB/s in maximum which is about 5 times of the deduplication systems used CPU only.

Cocktail: Exploiting Bartenders' Gestures for Mobile Interaction

Todays mobile devices are capable of creating and storing a large amount of multimedia data, but sharing such data with others remains challenging. Even when a user wishes to send a picture to a friend located next to the user, current wireless data transfer techniques usually demand several steps requiring inconvenient user involvement, such as manipulation of tiny buttons on a small touch screen. This paper presents Cocktail, a new gesture-based mobile interaction system that exploits gestures employed by bartenders for easy data sharing between co-located users. In our system, users can pour transfer multimedia data in their mobile devices to other friends devices in a manner akin to a bartender pouring a drink into a glass. Cocktail also provides an intuitive way of creating new content by mixing existing data with a shaking gesture. For example, users can make music videos with their favorite music and pictures in the mobile phone by selecting them and shaking the mobile phone. The authors have implemented a prototype of Cocktail with commercial smart phones and evaluated its usability via user studies.

MN-MATE: Elastic Resource Management of Manycores and a Hybrid Memory Hierarchy for a Cloud Node

Recent advent of manycore system increases needs for larger but faster memory hierarchy. Emerging next generation memories such as on-chip DRAM and nonvolatile memory (NVRAM) are promising candidates for replacement of DRAM-only main memory. Combined with the manycore trends, it gives an opportunity to rethink conventional resource management system with a memory hierarchy for a single cloud node. In an attempt to mitigate the energy and memory problems, we propose MN-MATE, an elastic resource management architecture for a single cloud node with manycores, on-chip DRAM, and large size of off-chip DRAM and NVRAM. In MN-MATE, the hypervisor places consolidated VMs and balances memory among them. Based on the monitored information about the allocated memory, a guest OS co-schedules tasks accessing different types of memory with complementary access intensity. Polymorphic management of DRAM hierarchy accelerates average memory access speed inside each guest OS. A guest OS reduces energy consumption with small performance loss based on the NVRAM-aware data placement policy and the hybrid page cache. A new lightweight kernel is developed to reduce the overhead from the guest OS for scientific applications. Experiment results show that our techniques in MN-MATE platform improve system performance and reduce energy consumption.

Dependence of polarization-mode dispersion penalties on decision threshold and receiver bandwidth

We measured the dependence of PMD-induced penalty on receiver characteristics in an optically preamplified receiver. The penalty is strongly dependent on the location of the decision point and on the electrical filter bandwidth of the receiver. We also confirmed that, for an optically preamplified receiver, RZ signals are more tolerant of first-order PMD than NRZ signals for typical receiver electrical bandwidths.

Credit-Based Runtime Placement of Virtual Machines on a Single NUMA System for QoS of Data Access Performance

While a NUMA system is being widely used as a target machine for virtualization, each data access request produced by a virtual machine (VM) on the NUMA system may have a different access time depending on not only remote access condition, but also shared resource contentions. Mainly due to this, each VM running on the NUMA system will have irregular data access performance at different times. Because existing hypervisors, such as KVM, VMware, and Xen, have yet to consider this, users of VMs cannot predict their data access performance or even recognize the data access performance they have experienced. In this paper, we propose a novel VM placement technique to resolve this issue pertaining to irregular data access performance of VMs running on the NUMA system. The hypervisor with our technique provides the illusion of a private memory subsystem to each VM, which guarantees the data access latency required by each VM on average. To enable this feature, we periodically evaluates the average data access latency of each VM using hardware performance monitoring units. After every evaluation, our Mcredit-based VM migration algorithm tries to migrate the VCPU or memory of the VM not meeting with its required data access latency to another node, giving the VM less data access latency. We implemented the prototype for KVM hypervisor on Linux 3.10.10. Experimental results show that, in the four-node NUMA system, our technique keeps the required data access performance levels of VMs running various workloads while it only consumes less than 1 percent of the cycles of a core and 0.3 percent of the system memory bandwidth.

Ablative microstructuring with plasma-based XUV lasers and efficient processing of materials by dual action of XUV/NIR–VIS ultrashort pulses

We report on a single-shot micropatterning of an organic polymer achieved by ablation with demagnifying projection using a plasma-based extreme ultraviolet (XUV) laser at 21 nm. A nickel mesh with a period of 100 μ m was 10×demagnified and imprinted on poly(methyl methacrylate) (PMMA) via direct ablation. This first demonstration of single-shot projection, single-step lithography illustrates the great potential of XUV lasers for the direct patterning of materials with a resolution scalable down to the submicrometer domain. In the second part, we present a novel experimental method for improving the efficiency of surface processing of various solids achieved by simultaneous action of XUV, obtained from high-order harmonic generation, and near-infrared (NIR)–VIS laser pulses. The NIR–VIS pulse interacts with free charge carriers produced by the energetic XUV photons, so that its absorption dramatically increases. Laser-induced periodic surface structures were effectively produced using this technique.

Impact of nonlinear crosstalk on optical PMD compensation

We have investigated the effects of nonlinear crosstalk on the polarisation mode dispersion (PMD) compensation in WDM systems. The results show that the first-order PMD compensation could degrade the performances of WDM systems due to the optical Kerr effect/DGD crosstalk.

Malfinder: Accelerated malware classification system through filtering on manycore system

Control flow matching methods have been utilized to detect malware variants. However, as the number of malware variants has soared, it has become harder and harder to detect all malware variants while maintaining high accuracy. Even though many researchers have proposed control flow matching methods, there is still a trade-off between accuracy and performance. To solve this trade-off, we designed Malfinder, a method based on approximate matching, which is accurate but slow. To overcome its low performance, we resolve its performance bottleneck and non-parallelism on three fronts: I-Filter for identical string matching, table division to exclude unnecessary comparisons with some malware and dynamic resource allocation for efficient parallelism. Our performance evaluation shows that the total performance improvement is 280.9 times.

R-Barrier: Rapid Barrier for Software RAID Cache Using Hints from Journaling Filesystem

While adopting cache in software RAID brings performance benefit, it can cause data loss at power failure which results in the broken filesystem consistency. Though I/O Barrier can be used to remove the consistency issue, it sacrifices the write performance of software RAID. To mitigate this tradeoff, we suggest R-Barrier for software RAID. The basic idea of R-Barrier is to avoid the synchronization of the entire cache space. Instead, based on given hints from the filesystem layer, R-Barrier makes software RAID cache select one of the following methods when the cache writes a certain write request into the disk: (1)Strictly ordered destaging for the write requests affecting the filesystem consistency, (2)Reordered destaging for the rest of write requests. We show that R-Barrier guarantees same filesystem consistency with I/O Barrier while the performance degradation of R-Barrier is less than that of I/O Barrier.