• International journal of advanced smart convergence
• /
• 제12권4호
• /
• pp.164-170
• /
• 2023

With the recent advances in AI (artificial intelligence) and HPC (high-performance computing) technologies, deep learning is proliferated in various domains of the 4th industrial revolution. As the workload volume of deep learning increasingly grows, analyzing the memory reference characteristics becomes important. In this article, we analyze the memory reference traces of deep learning workloads in comparison with traditional workloads specially focusing on read and write operations. Based on our analysis, we observe some unique characteristics of deep learning memory references that are quite different from traditional workloads. First, when comparing instruction and data references, instruction reference accounts for a little portion in deep learning workloads. Second, when comparing read and write, write reference accounts for a majority of memory references, which is also different from traditional workloads. Third, although write references are dominant, it exhibits low reference skewness compared to traditional workloads. Specifically, the skew factor of write references is small compared to traditional workloads. We expect that the analysis performed in this article will be helpful in efficiently designing memory management systems for deep learning workloads.

• Journal of Computing Science and Engineering
• /
• 제8권3호
• /
• pp.157-172
• /
• 2014

Contemporary embedded systems often use NAND flash memory instead of hard disks as their swap space of virtual memory. Since the read/write characteristics of NAND flash memory are very different from those of hard disks, an efficient page replacement algorithm is needed for this environment. Our analysis shows that temporal locality is dominant in virtual memory references but that is not the case for write references, when the read and write references are monitored separately. Based on this observation, we present a new page replacement algorithm that uses different strategies for read and write operations in predicting the re-reference likelihood of pages. For read operations, only temporal locality is used; but for write operations, both write frequency and temporal locality are used. The algorithm logically partitions the memory space into read and write areas to keep track of their reference patterns precisely, and then dynamically adjusts their size based on their reference patterns and I/O costs. Without requiring any external parameter to tune, the proposed algorithm outperforms CLOCK, CAR, and CFLRU by 20%-66%. It also supports optimized implementations for virtual memory systems.

• International journal of advanced smart convergence
• /
• 제12권1호
• /
• pp.70-75
• /
• 2023

With the recent proliferation of memory-intensive workloads such as deep learning, analyzing memory access characteristics for efficient memory management is becoming increasingly important. Since read and write operations in memory access have different characteristics, an efficient memory management policy should take into accountthe characteristics of thesetwo operationsseparately. Although some previous studies have considered the different characteristics of reads and writes, they require a modified hardware architecture supporting read bits and write bits. Unlike previous approaches, we propose a software-based management policy under the existing memory architecture for considering read/write characteristics. The proposed policy logically partitions memory space into the read/write area and the write area by making use of reference bits and dirty bits provided in modern paging systems. Simulation experiments with memory access traces show that our approach performs better than the CLOCK algorithm by 23% on average, and the effect is similar to the previous policy with hardware support.

• JSTS:Journal of Semiconductor Technology and Science
• /
• 제16권1호
• /
• pp.23-30
• /
• 2016

We report a STT-MRAM write-scheme, in which the length of the write-pulse is determined dynamically by sensing the status of MTJ cells. The proposed scheme can reduce the power consumption by eliminating unnecessary writing current after the switching has occurred. We also propose a reference cell design, which is optimized for the use in write-circuits. The performance of the proposed circuit was verified by SPICE level simulations of the circuit implemented in a $0.13{\mu}m$ CMOS process.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• 제23권1호
• /
• pp.71-76
• /
• 2023

Recently, a memory management strategy that utilizes read and write references separately is attracting attention. This is due to the emergence of new storage media with asymmetric read/write latencies and different read/write access characteristics of software. Existing research assumes that operating systems can differentiate between read/write references that occur on each memory page, but most memory architectures do not support a way to distinguish them. Unlike previous studies, this paper proposes a software method that reflects the read/write characteristics of page references by utilizing the reference and modified bits of each page. Simulations show that the proposed policy has almost similar effects to existing studies with hardware support.

• Journal of KIISE:Computer Systems and Theory
• /
• 제36권6호
• /
• pp.543-556
• /
• 2009

Recently, NAND flash memory is being used as the swap device of virtual memory as well as the file storage of mobile systems. Since temporal locality is dominant in page references of virtual memory, LRU and its approximated CLOCK algorithms are widely used. However, cost of a write operation in flash memory is much larger than that of a read operation, and thus a page replacement algorithm should consider this factor. This paper analyzes virtual memory read/write reference patterns individually, and observes the ranking inversion problem of temporal locality in write references which is not observed in read references. With this observation, we present a new page replacement algorithm considering write frequency as well as temporal locality in estimating write reference behaviors. This new algorithm dynamically allocates memory space to read/write operations based on their reference patterns and I/O costs. Though the algorithm has no external parameter to tune, it supports optimized implementations for virtual memory systems, and also performs 20-66% better than CLOCK, CAR, and CFLRU algorithms.

• Journal of the Korea Society of Computer and Information
• /
• 제25권2호
• /
• pp.1-10
• /
• 2020

In this paper, we propose an efficient disk buffer replacement policy which improves hit ratio and reduces writing operations of flash based storages. The flash based storage has many advantages, including a small form factor, non-volatility and high reliability, but there are problems caused by own limitations, like not-in-place update, short life cycle and asymmetric I/O latencies. To redeem these problems, this paper proposes the write weighted probability of reference(2WPR) policy. 2WPR policy predicts re-referencing probability and calculates localities of each page. Furthermore, by weighting write operations to every pages, 2WPR can reduce write operations to flash based storage. In addition, we can improve the performance with higher hit ratio and reduce the number of write operations and consequently shorten the latencies of each operation. The results show that our policy provides improvements of up to 10% for the hit ratio with the reduction of up to 5% for the flash writing operation compared with other policies.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• 제21권6호
• /
• pp.65-70
• /
• 2021

Recently, as the number of mobile apps rapidly increases, the memory size of smartphones keeps increasing. Smartphone memory consists of DRAM and as it is a volatile medium, continuous refresh operations for all cells should be performed to maintain the contents. Thus, the power consumption of memory increases in proportion to the DRAM size of the system. There are attempts to configure the memory system with low-power non-volatile memory instead of DRAM to reduce the power consumption of smartphones. However, non-volatile memory has weaknesses in write operations, so analysis of write behaviors is a prerequisite to realize this in practical systems. In this paper, we extract memory reference traces of mobile apps and analyze their characteristics specially focusing on write operations. The results of this paper will be helpful in the design of memory management systems consisting of non-volatile memory in future smartphones.

• Journal of the Institute of Electronics Engineers of Korea SC
• /
• 제39권2호
• /
• pp.132-139
• /
• 2002

This paper describes a supply voltage detector and SRAM write driver circuit which dissipates small power. The supply voltage detector generates high signal when supply voltage is higher than reference voltage, but low signal when supply voltage is lower than reference voltage. The write driver utilizes two same-sized drivers to reduce operating current in the write cycle. In the case of lower supply voltage comparing to Vcc, both drivers are active the same as conventional write driver, while in the case of high Vcc only one of two drivers are active so as to deliver the half of the current. As a result of simulation using 0.6${\mu}{\textrm}{m}$ 3.3v/5v, CMOS model parameter, the proposed SRAM scheme shows a 22.6% power reduction and 12.7% PDP reduction at Vcc=3.3V, compared to the conventional one.

• The Transactions of the Korea Information Processing Society
• /
• 제7권3호
• /
• pp.776-787
• /
• 2000

Write requests which stored in disk cache of the RAID5 controller should be destaged to disk arrays according to the destage algorithm. As the response performance of the parallel IO request is being hit by the effect of the destage, several destage algorithms have been studied to enhance the performance of he RAID5 controller. Among the destage algorithms, the load balancing destage algorithm has better performance than other destage algorithms when system load is highly increased. But the load balancing destage algorithm gives priority to load balance among the disks in disk arrays, therefore, when some disks are affected by the very heavy system load caused by small data requests, the load balancing destage algorithm cannot enhance the performance of parallel IO requests effectively since it makes effort to maintain the load balance without the benefit of the locality of the write requests. This paper proposes a new RAID5 controller that applied reference-load balancing destage algorithm which decides the destage priority based on the reference history and load distribution of the disks. The simulation results show that RAID5 controller with the reference-load balancing destage algorithm has better performance than previous load balancing destage algorithm.