Spark is a distributed data processing framework based on memory.Memory allocation is a focus question of Spark research.A good memory allocation scheme can effectively improve the efficiency of task execution and mem...Spark is a distributed data processing framework based on memory.Memory allocation is a focus question of Spark research.A good memory allocation scheme can effectively improve the efficiency of task execution and memory resource utilization of the Spark.Aiming at the memory allocation problem in the Spark2.x version,this paper optimizes the memory allocation strategy by analyzing the Spark memory model,the existing cache replacement algorithms and the memory allocation methods,which is on the basis of minimizing the storage area and allocating the execution area according to the demand.It mainly including two parts:cache replacement optimization and memory allocation optimization.Firstly,in the storage area,the cache replacement algorithm is optimized according to the characteristics of RDD Partition,which is combined with PCA dimension.In this section,the four features of RDD Partition are selected.When the RDD cache is replaced,only two most important features are selected by PCA dimension reduction method each time,thereby ensuring the generalization of the cache replacement strategy.Secondly,the memory allocation strategy of the execution area is optimized according to the memory requirement of Task and the memory space of storage area.In this paper,a series of experiments in Spark on Yarn mode are carried out to verify the effectiveness of the optimization algorithm and improve the cluster performance.展开更多
Caching is an important technique to enhance the efficiency of query processing. Unfortunately, traditional caching mechanisms are not efficient for deep Web because of storage space and dynamic maintenance limitation...Caching is an important technique to enhance the efficiency of query processing. Unfortunately, traditional caching mechanisms are not efficient for deep Web because of storage space and dynamic maintenance limitations. In this paper, we present on providing a cache mechanism based on Top-K data source （KDS-CM） instead of result records for deep Web query. By integrating techniques from IR and Top-K, a data reorganization strategy is presented to model KDS-CM. Also some measures about cache management and optimization are proposed to improve the performances of cache effectively. Experimental results show the benefits of KDS-CM in execution cost and dynamic maintenance when compared with various alternate strategies.展开更多
Nowadays,household NAS service provides consumer device users a convenient way to extend their storage space through WLAN.As the core of NAS service,distributed file system’s objective is to make the whole system pro...Nowadays,household NAS service provides consumer device users a convenient way to extend their storage space through WLAN.As the core of NAS service,distributed file system’s objective is to make the whole system provide a good user experience.However,it is normal for consumer devices to work under unstable network conditions,which can significantly degrade the user experience.Moreover,user behavior is getting more complicated,which means traditional mainstream optimization techniques have little effect in this case.To improve user experience,this paper proposes several practical optimization solutions for a lightweight distributed file system.First,a set of client-side cache optimization schemes,including swap-based persistent caching and cross-device cache prefetching with the Markov method,are proposed to reduce remote access latency.Second,a task-aware write-back scheduling scheme is proposed to enhance the cache synchronization efficiency.Finally,a simple protocol based on multiple readers and single writer for multi-device access control is proposed.Experiments on real devices show that the average access latency can be reduced by 29.7%with swap-based client-side persistent caching.Cross-device prefetching reduces around 33%access latency in the best case.Average cache synchronization latency is reduced by 13.7%and the worst synchronization latency is reduced by 63.7%with write-back scheduling.Multi-device access protocol induced negligible overhead but works effectively on controlling concurrent accesses.展开更多
Due to flexible deployment,unmanned aerial vehicle(UAV)mounted aerial access points are capable of expanding the coverage capabilities of existing terrestrial base stations(TBSs).Different from TBSs,however,UAV access...Due to flexible deployment,unmanned aerial vehicle(UAV)mounted aerial access points are capable of expanding the coverage capabilities of existing terrestrial base stations(TBSs).Different from TBSs,however,UAV access points(UAPs)are of high mobility in horizontal and vertical dimensions,which may deteriorate the coverage performance.Worsestill,the mobility of UAPs would as well increase the pressure of wireless backhaul.In this light,we investigate the performance of the cache-enabled UAV communications network(CUCN)in terms of network spatial throughput(ST)by analyzing the line of sight(LoS)connections and non-line of sight(NLoS)connections.It is found that the network ST is exponentially decreased with the square of UAP altitude.Furthermore,contrary to intuition,a large cache size may deteriorate the network ST when UAPs are over-deployed.The reason is that a large cache size increases the hit probability,which may increase the activation of UAPs and consequently result in complicated interference.Aiming to maximize the network ST,we optimize the cache strategy under limited backhaul.Remarkably,the results show that network ST could be substantially improved by the optimized cache strategy and the performance degeneration brought by UAP high mobility could be even eliminated especially when the UAP altitude is high.展开更多
In recent years,the exponential growth in Large Language Model(LLM)parameter sizes has significantly increased computational complexity,with inference latency emerging as a prominent challenge.The primary bottleneck l...In recent years,the exponential growth in Large Language Model(LLM)parameter sizes has significantly increased computational complexity,with inference latency emerging as a prominent challenge.The primary bottleneck lies in the token-by-token prediction process during autoregressive decoding,resulting in substantial delays.Therefore,enhancing decoding efficiency while maintaining accuracy has become a critical research objective.This paper proposes an Adaptive Parallel Layer-Skipping Speculative Decoding(APLS)method,which leverages speculative decoding techniques by employing a Small-Scale Model(SSM)for preliminary inference and validating the predictions using the original LLM.This approach effectively balances the high precision of LLMs with the efficiency of SSMs.Notably,our SSM does not require additional training but is instead derived through a simplification of the original large-scale model.By incorporating parallelization and a layer-skipping structure,the inference process dynamically bypasses certain redundant transformation layers,significantly improving GPU utilization and inference speed without compromising performance.Furthermore,to address challenges such as window size limitations and memory fragmentation in long-text processing,this paper introduces progressive layer reduction and key-value cache deletion techniques to further optimize the performance of SSMs.Experimental results demonstrate that the proposed method achieves a 2.51×improvement in efficiency during autoregressive decoding.As this approach eliminates the need for additional training of SSM,it offers a significant competitive advantage in high-cost model compression environments.展开更多
文摘Spark is a distributed data processing framework based on memory.Memory allocation is a focus question of Spark research.A good memory allocation scheme can effectively improve the efficiency of task execution and memory resource utilization of the Spark.Aiming at the memory allocation problem in the Spark2.x version,this paper optimizes the memory allocation strategy by analyzing the Spark memory model,the existing cache replacement algorithms and the memory allocation methods,which is on the basis of minimizing the storage area and allocating the execution area according to the demand.It mainly including two parts:cache replacement optimization and memory allocation optimization.Firstly,in the storage area,the cache replacement algorithm is optimized according to the characteristics of RDD Partition,which is combined with PCA dimension.In this section,the four features of RDD Partition are selected.When the RDD cache is replaced,only two most important features are selected by PCA dimension reduction method each time,thereby ensuring the generalization of the cache replacement strategy.Secondly,the memory allocation strategy of the execution area is optimized according to the memory requirement of Task and the memory space of storage area.In this paper,a series of experiments in Spark on Yarn mode are carried out to verify the effectiveness of the optimization algorithm and improve the cluster performance.
基金Supported by the National Natural Science Foundation of China (60673139, 60473073, 60573090)
文摘Caching is an important technique to enhance the efficiency of query processing. Unfortunately, traditional caching mechanisms are not efficient for deep Web because of storage space and dynamic maintenance limitations. In this paper, we present on providing a cache mechanism based on Top-K data source （KDS-CM） instead of result records for deep Web query. By integrating techniques from IR and Top-K, a data reorganization strategy is presented to model KDS-CM. Also some measures about cache management and optimization are proposed to improve the performances of cache effectively. Experimental results show the benefits of KDS-CM in execution cost and dynamic maintenance when compared with various alternate strategies.
基金supported by the NSFC 62141213,Shanghai Science and Technology Project(22QA1403300)the Open Project Program of Wuhan National Laboratory for Optoelectronics NO.2019WNLOKF008.
文摘Nowadays,household NAS service provides consumer device users a convenient way to extend their storage space through WLAN.As the core of NAS service,distributed file system’s objective is to make the whole system provide a good user experience.However,it is normal for consumer devices to work under unstable network conditions,which can significantly degrade the user experience.Moreover,user behavior is getting more complicated,which means traditional mainstream optimization techniques have little effect in this case.To improve user experience,this paper proposes several practical optimization solutions for a lightweight distributed file system.First,a set of client-side cache optimization schemes,including swap-based persistent caching and cross-device cache prefetching with the Markov method,are proposed to reduce remote access latency.Second,a task-aware write-back scheduling scheme is proposed to enhance the cache synchronization efficiency.Finally,a simple protocol based on multiple readers and single writer for multi-device access control is proposed.Experiments on real devices show that the average access latency can be reduced by 29.7%with swap-based client-side persistent caching.Cross-device prefetching reduces around 33%access latency in the best case.Average cache synchronization latency is reduced by 13.7%and the worst synchronization latency is reduced by 63.7%with write-back scheduling.Multi-device access protocol induced negligible overhead but works effectively on controlling concurrent accesses.
基金supported in part by National Key Research and Development Program of China (Grant No. 2020YFB1807001)in part by Natural Science Foundation of China (Grant No. 62171344, 62121001, 61725103, 61931005)+1 种基金in part by Young Elite Scientists Sponsorship Program by CASTin part by Key Industry Innovation Chain of Shaanxi (Grant No. 2022ZDLGY05-01, 2022ZDLGY05-06)
文摘Due to flexible deployment,unmanned aerial vehicle(UAV)mounted aerial access points are capable of expanding the coverage capabilities of existing terrestrial base stations(TBSs).Different from TBSs,however,UAV access points(UAPs)are of high mobility in horizontal and vertical dimensions,which may deteriorate the coverage performance.Worsestill,the mobility of UAPs would as well increase the pressure of wireless backhaul.In this light,we investigate the performance of the cache-enabled UAV communications network(CUCN)in terms of network spatial throughput(ST)by analyzing the line of sight(LoS)connections and non-line of sight(NLoS)connections.It is found that the network ST is exponentially decreased with the square of UAP altitude.Furthermore,contrary to intuition,a large cache size may deteriorate the network ST when UAPs are over-deployed.The reason is that a large cache size increases the hit probability,which may increase the activation of UAPs and consequently result in complicated interference.Aiming to maximize the network ST,we optimize the cache strategy under limited backhaul.Remarkably,the results show that network ST could be substantially improved by the optimized cache strategy and the performance degeneration brought by UAP high mobility could be even eliminated especially when the UAP altitude is high.
基金the National Natural Science Foundation of China under Grant 62404256Jiangsu Provincial Science and Technology Major Special Project under Grant BG2024032+1 种基金the Key Project of Shenzhen Basic Research Program under Grant JCYJ20241206180301003the High-performance Computing Public Platform(Shenzhen Campus)of Sun Yat-sen University.
文摘In recent years,the exponential growth in Large Language Model(LLM)parameter sizes has significantly increased computational complexity,with inference latency emerging as a prominent challenge.The primary bottleneck lies in the token-by-token prediction process during autoregressive decoding,resulting in substantial delays.Therefore,enhancing decoding efficiency while maintaining accuracy has become a critical research objective.This paper proposes an Adaptive Parallel Layer-Skipping Speculative Decoding(APLS)method,which leverages speculative decoding techniques by employing a Small-Scale Model(SSM)for preliminary inference and validating the predictions using the original LLM.This approach effectively balances the high precision of LLMs with the efficiency of SSMs.Notably,our SSM does not require additional training but is instead derived through a simplification of the original large-scale model.By incorporating parallelization and a layer-skipping structure,the inference process dynamically bypasses certain redundant transformation layers,significantly improving GPU utilization and inference speed without compromising performance.Furthermore,to address challenges such as window size limitations and memory fragmentation in long-text processing,this paper introduces progressive layer reduction and key-value cache deletion techniques to further optimize the performance of SSMs.Experimental results demonstrate that the proposed method achieves a 2.51×improvement in efficiency during autoregressive decoding.As this approach eliminates the need for additional training of SSM,it offers a significant competitive advantage in high-cost model compression environments.
