분류 전체보기30 [SC'21] DeltaFS: A Scalable No-Ground-Truth Filesystem For Massively-Parallel Computing [SC'21] DeltaFS: A Scalable No-Ground-Truth Filesystem For Massively-Parallel Computing 1. Motivations Global synchronization Today's filesystem clients tend to synchronize too frequently with their servers. The inadequacy of the current state-of-art Today's filesystems map all application jobs to a single filesystem namespace. Filesystem metadata performance is limited by the amount.. 2022. 9. 6. [HotStorage'22] Lifetime-leveling LSM-tree compaction for ZNS SSD [HotStorage'22] Lifetime-leveling LSM-tree compaction for ZNS SSD 1. Motivations Problems from compaction in LSM-based KV stores with ZNS SSD Space amplification (Long-lived SSTs) Write amplification (Short-lived SSTs) 2. LL (Lifetime leveling) compaction algorithm Key principles Allocates dedicated zones for each level (For avoiding long-lived SSTs) Each compaction must involve all the lowe.. 2022. 7. 25. ATC & OSDI'22 7/11 KEY NOTE: Surprize-Inspired Networking what will be surplus? Edge capacity High resolution sensors how will future apps be different? they will be proactive E2C (Edge-to-Cloud) [ATC] ZNSwap: un-Block your Swap Motivation Swap on traditional SSDs certain performance anomalies Flash SSDs ZNSwap : new sw33ap subsystem for ZNS [OSDI] BlockFlex HarvestVMs: Promising for Improved Resource Utiliza.. 2022. 7. 18. [FAST'19] SLM-DB: Single-Level Key-Value Store with Persistent Memory [FAST'19] SLM-DB: Single-Level Key-Value Store with Persistent Memory 1. Motivations LSM-tree based KV stores are optimized to support write intensive workloads It has high R/W amplification and low read performance Recently, typical workloads has changed to have similar proportions of R/W KV stores need to be optimized for both read and write workloads 2. Optimization point Improve read perfoma.. 2022. 7. 18. [HotStorage'22] Compaction-Aware Zone Allocation for LSM based Key-Value Store on ZNS SSDs [HotStorage'22] Compaction-Aware Zone Allocation for LSM based Key-Value Store on ZNS SSDs 1. Motivations ZNS SSD Space management is moved from device to host Thus, application must perform zone cleaning Write amplification: When cleaning a zone, if valid data remains, this data should be copied to other zone To reduce WA, data with the same lifetime should be written in the same zone LIZA(LIfe.. 2022. 7. 18. [ATC'18] Redesigning LSMs for Nonvolatile Memory with NoveLSM [ATC'18] Redesigning LSMs for Nonvolatile Memory with NoveLSM 1. Critical problems on LSM when we use NVM Different in-memory and storage(persistent) form of the data High (de)serialization cost Only in-memory data can be changed. storage data is immutable. Limited memory capacity leads to frequent compaction, which increases stall time Becuase memory is volatile, updates must be logged. This al.. 2022. 7. 4. [ATC'20] MatrixKV: Reducing Write Stalls and Write Amplification in LSM-tree Based KV Stores with Matrix Container in NVM [ATC'20] MatrixKV: Reducing Write Stalls and Write Amplification in LSM-tree Based KV Stores with Matrix Container in NVM 1. Challenges on LSM-tree based KV stores Write stalls LSM구조에서 L0레벨은 DRAM에서 flush된 Memtable이 그대로 저장되기 때문에 정렬이 되어 있지 않다. Compaction은 merge sort 방식으로 하는데, L0이 정렬된 상태가 아니기 때문에 여기서 오는 overheadr가 크다. Write amplification WA = n*AF (n: # of levels, AF: Amplification factor) 2. Previ.. 2022. 6. 30. [ASPLOS'19] Nimble Page Management for Tiered Memory Systems [ASPLOS'19] Nimble Page Management for Tiered Memory Systems Ref: https://dl.acm.org/doi/10.1145/3297858.3304024 1. Introduction 현대의 processing, memory system에서 hetrogeneity가 많아지는 추세 효율적인 자원의 관리가 필요 hot data는 fast memory, cold data는 slow memory에 배치 이 hot, cold data migration의 오버헤드가 너무 커서 hardware의 banwidth에 한참 미치지 못한다 이 논문에서 제시하는 4가지 optimziation Huge page migration Parallelized data copy Concur.. 2022. 1. 14. [SOSP'21] HeMem: Scalable Tiered Memory Management for Big Data Applications and Real NVM [SOSP'21] HeMem: Scalable Tiered Memory Management for Big Data Applications and Real NVM Ref: https://dl.acm.org/doi/10.1145/3477132.3483550 2.Background 2.1 Application Memory Demands 현대의 application들은 data-intensive ex) web, machine learning, in-memory DB... Requirements large memory capacity, bandwidth low memory access tail-latency 위에서 언급한 application들의 memory lifetime 특성은 allocation size와 .. 2022. 1. 5. RocksDB RocksDB Ref: https://github.com/facebook/rocksdb/wiki/RocksDB-Overview 1. Overview persistent key-value store Key, value들은 arbitrarily-sized byte stream user-specified comparator function으로 key가 ordered됨 pure memory, flash memory, hdd, remote storage 등 다양한 production env에 적용이 될 수 있다,. 2. Assumptions and Goals Performance design point : it should be performant for fast storage and for server work.. 2022. 1. 5. wsl에서 Pytorch Cuda버전 설치하기(+ VSCode) * OS: Windows 11 - wsl2 * 윈도우 10은 wsl에서 host gpu에 접근이 불가능 * Wsl 설치 방법 : https://webdir.tistory.com/541 * Wsl Cuda Driver 설치 * https://developer.nvidia.com/cuda/wsl * pytorch_gpu 설치 * Anaconda 설치 * https://www.anaconda.com/products/individual 에서 Linux 버전으로 받은 후 sh파일 실행 * CUDA version 체크 $nvcc --version * https://pytorch.kr/get-started/revious-versions/ 에서 cuda에 맞는 pytorch 설치 * 설치 커맨드를 입력하면 저렇게 세부.. 2021. 11. 3. [Chapter 5] CPU scheduling [Chapter 5] CPU Scheduling 5.1 Basic Concepts CPU scheduling의 목적 : Maximize CPU utilization! CPU-I/O Burst Cycle Process execution은 CPU execution과 I/O wait으로 나누어진다. 프로그램은 실행 특성에 따라 구분할 수 있다 CPU-bound program : 적은 수의 long CPU burst I/O보다는 직접적인 computation을 많이 하기 때문에 CPU 사용에 시간을 오랫동안 씀 I/O-bound program : 많은 수의 short CPU burst I/O가 많기 때문에 짧은 시간동안 CPU를 이용하고 I/O대기에 쓰는 시간이 많음 이러한 구분이 CPU scheduling a.. 2021. 7. 27. 이전 1 2 3 다음
