计算所科研实践随笔
被淹没在论文海里的两个星期。
早上7:10分起床,草草洗漱,7:30出发,开始漫长的1小时通勤。从地铁站的安检口起,队便排的极长,让人看得头皮发麻。下到了轨道旁稍好,但每趟呼啸而来的地铁里都是满满当当,常常要等2,3趟我才能有幸站上去。我从不奢望座位。在等一趟不那么满的地铁的过程中,有时很满的地铁来了,门易开人墙难开,门快关时有些赶得急的人就踏上地铁背过身来拼命地往后压,最后强行挤出了个位置来,我在外面看得见那努力的表情。
虽说这行为不大雅观,素质上可能偏向于负面,但我脑子里总是闪现那句“我用尽了全力,过着平凡的一生”。
地铁门关,轻灵驰走,满载着刚开始新的一天却已有些疲惫的人们。
到了计算所后在负一楼的健康餐吧吃早饭。这是艰难的两个星期里比较安慰我的两样之一。两样一个是计算所的伙食,一个是友善的师兄。
这两周导师给我的任务是RDMA调研。初步了解RDMA技术,找出RDMA小领域里50篇好的论文并每篇写概述(写不出来就翻译摘要),最后写一篇总结与感想。
看论文的过程真的是很难受的。我才刚大一结束,很多基础知识都不知道,论文里使用的大量术语我基本都知形不知义。于是不停wiki,常常wiki里对术语的解释本身便是由术语写就的,我便只好继续wiki。如此树形展开,搜好几层我才能大致明白某个术语的含义。不过很多时候我都直接叨扰了友善的师兄。
时间与我支付宝里的钱一同流逝着,到了后面,我越来越浮躁。对待每一篇论文的态度越来越敷衍。这一点友善的师兄在最后为导师代笔的评语了也有提到。哎,是心不大静。
努力前行。
计算所科研实践报告
在调研过程中使用的搜集信息方式:
- 中国知网搜索:搜到的论文质量较差,舍弃。故没有产生作用。
- RDMA技术标准制定方官网https://www.openfabrics.org/:用于了解技术原理。
- 一个有关RDMA的博客http://www.rdmamojo.com/:用于了解技术原理。
- *:使用频繁,主要用于查看术语定义。
- Google scholar搜索:主力,大部分论文由此获取。
- 查看论文作者的其他论文:查阅过程中发现Dhabaleswar K.DK Panda作者的名字常常出现,于是查了一下该作者的其他论文。
- 问师兄:非常频繁,帮助很大。
- 后来在中国科学技术大学罗昭锋老师放在科学网的一篇博文《引文分析软件histcite简介》的启发下利用histcite软件分析了Web of Science核心合集数据库中与RDMA相关的约500篇论文,并进行可视化
RDMA技术简介
利用RDMA的特殊网卡,在不需要操作系统的干预下,交换虚拟地址,内存钥匙,数据等信息,从而实现远程直接访问内存。由于1零拷贝2无CPU干预3所有协议直接写入channel adapter,最终可以降低延迟,CPU占用率,提高带宽利用率。
现主要有三种产品:InfiniBand,IWARP(internet Wide Area RDMA Protocol),RoCE(RDMA over Converged Ethernet)
RDMA与TCP的比较
成本上RDMA较贵:
10G普通网卡:Mellanox MCX341A-XCCN ConnectX-3 10Gigabit Ethernet Card
$170.00
Infiniband交换机:Mellanox MCS7510 66Tb/s, 324-port EDR Infiniband chassis switch, includes 12 fans and 6 power supplies (N+N), RoHS R6
$66,950.00
infiniband卡:Mellanox MCX456A-ECAT ConnectX VPI Infiniband Host Bus Adapter
$1,575.00
论文分类:
1.对RDMA读操作的研究
分析RDMA读操作不同情况下的利弊
Dragojevic A, Narayanan D, Castro M. RDMA Reads: To Use or Not to Use?[J]. IEEE Data Eng. Bull., 2017, 40(1): 3-14.
键值对应用中未使用读操作的方案
Kalia A, Kaminsky M, Andersen D G. Using RDMA efficiently for key-value services[C]//ACM SIGCOMM Computer Communication Review. ACM, 2014, 44(4): 295-306.
键值对应用中使用了读操作但做出改进的方案
C. Mitchell, Y. Geng, and J. Li. Using One-Sided RDMA Reads to Build a Fast,
CPU-Efficient Key-Value Store. In USENIX ATC, 2013.
分析发现在广域网上RDMA读操作性能较差
Yu W, Rao N S V, Wyckoff P, et al. Performance of RDMA-capable storage protocols on wide-area network[C]//Petascale Data Storage Workshop, 2008. PDSW'08. 3rd. IEEE, 2008: 1-5.
MPI应用中将Rendezvous Protocol的实现由写操作实现改为使用读操作实现。
Sur S, Jin H W, Chai L, et al. RDMA read based rendezvous protocol for MPI over InfiniBand: design alternatives and benefits[C]//Proceedings of the eleventh ACM SIGPLAN symposium on Principles and practice of parallel programming. ACM, 2006: 32-39.
提出的优化方案中避免RDMA单边操作(RDMA读/写)
Kalia A, Kaminsky M, Andersen D G. FaSST: Fast, Scalable and Simple Distributed Transactions with Two-Sided (RDMA) Datagram RPCs[C]//OSDI. 2016: 185-201.
这篇研究RDMA隐藏成本的分析中也有提及:
Author(s): Frey PW (Frey, Philip W.); Alonso G (Alonso, Gustavo)
Title: Minimizing the Hidden Cost of RDMA
Source: 2009 29TH IEEE INTERNATIONAL CONFERENCE ON DISTRIBUTED COMPUTING SYSTEMS : 553-+
Date: 2009
DOI: 10.1109/ICDCS.2009.32
2.对RDMA虚拟化,软件化的研究
这个方向在2011年被提出来一次,作者们研究设计了一个软件RDMA栈,并试图应用在商业云上。
Trivedi A, Metzler B, Stuedi P. A case for RDMA in clouds: turning supercomputer networking into commodity[C]//Proceedings of the Second Asia-Pacific Workshop on Systems. ACM, 2011: 17.
最近又有两篇论文关注这方面的问题:
虚拟化RDMA(新,2017年论文)
Fan S, Chen F, Rauchfuss H, et al. Towards a Lightweight RDMA Para-Virtualization for HPC[C]//COSH/VisorHPC@ HiPEAC. 2017: 39-44.
软件化RDMA。(新,2017年论文)
Mao Miao, Fengyuan Ren, Xiaohui Luo, Jing Xie, Qingkai Meng, and Wenxue Cheng. 2017. SoftRDMA: Rekindling High Performance Software RDMA over Commodity Ethernet. In Proceedings of the First Asia-Pacific Workshop on Networking (APNet'17). ACM, New York, NY, USA, 43-49. DOI: https://doi.org/10.1145/3106989.3106995
3.对RDMA内存管理的研究
一个提高RDMA内存管理效率的方案(一个计算所老师的论文)
Ou L, He X, Han J. An efficient design for fast memory registration in RDMA[J]. Journal of Network and Computer Applications, 2009, 32(3): 642-651.
在MPI应用中一个提高RDMA内存管理效率的方案
Mamidala A, Vishnu A, Panda D. Efficient shared memory and RDMA based design for MPI_Allgather over infiniband[J]. Recent Advances in Parallel Virtual Machine and Message Passing Interface, 2006: 66-75.
设计了一个新协议来Overlap内存管理的开销,跟单边操作(读/写)也有关系
Woodall T, Shipman G, Bosilca G, et al. High performance RDMA protocols in HPC[J]. Recent Advances in Parallel Virtual Machine and Message Passing Interface, 2006: 76-85.
4.将RDMA应用于GPU的研究
以下四篇论文成如上图关系:
第一篇首先提出使用GPUDirect RDMA来实现 Inter-node MPI Communication:
Potluri S, Hamidouche K, Venkatesh A, et al. Efficient inter-node MPI communication using GPUDirect RDMA for InfiniBand clusters with NVIDIA GPUs[C]//Parallel Processing (ICPP), 2013 42nd International Conference on. IEEE, 2013: 80-89.
有3篇跟进:
Wang H, Potluri S, Bureddy D, et al. GPU-aware MPI on RDMA-enabled clusters: Design, implementation and evaluation[J]. IEEE Transactions on Parallel and Distributed Systems, 2014, 25(10): 2595-2605.
Hamidouche K, Venkatesh A, Awan A A, et al. Exploiting GPUDirect RDMA in Designing High Performance OpenSHMEM for NVIDIA GPU Clusters[C]//Cluster Computing (CLUSTER), 2015 IEEE International Conference on. IEEE, 2015: 78-87.
Banerjee D S, Hamidouche K, Panda D K. Designing high performance communication runtime for GPU managed memory: early experiences[C]//Proceedings of the 9th Annual Workshop on General Purpose Processing using Graphics Processing Unit. ACM, 2016: 82-91.
5.将RDMA应用于键值对存储的研究
这篇主要关注安全性问题:
Yang M, Yu S, Yu R, et al. InnerCache: A Tactful Cache Mechanism for RDMA-Based Key-Value Store[C]//Web Services (ICWS), 2016 IEEE International Conference on. IEEE, 2016: 646-649.
提高性能(在RDMA读操作部分中介绍过)
Kalia A, Kaminsky M, Andersen D G. Using RDMA efficiently for key-value services[C]//ACM SIGCOMM Computer Communication Review. ACM, 2014, 44(4): 295-306.
提高性能
C. Mitchell, Y. Geng, and J. Li. Using One-Sided RDMA Reads to Build a Fast,
CPU-Efficient Key-Value Store. In USENIX ATC, 2013.
这个主要是关注基于RDMA的键值对应用中固态硬盘和内存混合存储的问题
Shankar D, Lu X, Islam N, et al. High-Performance Hybrid Key-Value Store on Modern Clusters with RDMA Interconnects and SSDs: Non-blocking Extensions, Designs, and Benefits[C]//Parallel and Distributed Processing Symposium, 2016 IEEE International. IEEE, 2016: 393-402.
6.将RDMA应用于关系数据库的研究
共享内存避免读硬盘,从而提升性能
Li F, Das S, Syamala M, et al. Accelerating relational databases by leveraging remote memory and RDMA[C]//Proceedings of the 2016 International Conference on Management of Data. ACM, 2016: 355-370.
引入MySQL
Shankar D, Lu X, Jose J, et al. Can RDMA benefit online data processing workloads on memcached and MySQL?[C]//Performance Analysis of Systems and Software (ISPASS), 2015 IEEE International Symposium on. IEEE, 2015: 159-160.
加速数据库join操作
Barthels C, Loesing S, Alonso G, et al. Rack-scale in-memory join processing using RDMA[C]//Proceedings of the 2015 ACM SIGMOD International Conference on Management of Data. ACM, 2015: 1463-1475.
两篇都是利用RDMA和HTM设计新的Transactions progress system ,作者也一样
Wei X, Shi J, Chen Y, et al. Fast in-memory transaction processing using RDMA and HTM[C]//Proceedings of the 25th Symposium on Operating Systems Principles. ACM, 2015: 87-104.
Chen Y, Wei X, Shi J, et al. Fast and general distributed transactions using RDMA and HTM[C]//Proceedings of the Eleventh European Conference on Computer Systems. ACM, 2016: 26.
加速分布式Transactions时避免了RDMA单边操作
Kalia A, Kaminsky M, Andersen D G. FaSST: Fast, Scalable and Simple Distributed Transactions with Two-Sided (RDMA) Datagram RPCs[C]//OSDI. 2016: 185-201.
7.将RDMA应用于大数据系统的研究
将RDMA引入Spark提高性能
Lu X, Rahman M W U, Islam N, et al. Accelerating spark with RDMA for big data processing: Early experiences[C]//High-performance interconnects (HOTI), 2014 IEEE 22nd annual symposium on. IEEE, 2014: 9-16.
与上篇工作类似
Yan X, Wong B, Choy S. R3S: RDMA-based RDD Remote Storage for Spark[C]//Proceedings of the 15th International Workshop on Adaptive and Reflective Middleware. ACM, 2016: 4.
用RDMA解决MapReduce,Spark和HBase的主要存储引擎HDFS的I/O瓶颈,从而提升性能
Islam N S, Rahman M W, Jose J, et al. High performance RDMA-based design of HDFS over InfiniBand[C]//Proceedings of the International Conference on High Performance Computing, Networking, Storage and Analysis. IEEE Computer Society Press, 2012: 35.
这篇与上篇工作类似
Woodall T, Shipman G, Bosilca G, et al. High performance RDMA protocols in HPC[J]. Recent Advances in Parallel Virtual Machine and Message Passing Interface, 2006: 76-85.
引入Hadoop后的性能分析
Lu X, Islam N S, Wasi-Ur-Rahman M, et al. High-performance design of Hadoop RPC with RDMA over InfiniBand[C]//Parallel Processing (ICPP), 2013 42nd International Conference on. IEEE, 2013: 641-650.
设计了一个插件库帮助Hadoop使用RDMA
Wang Y, Xu C, Li X, et al. JVM-bypass for efficient Hadoop shuffling[C]//Parallel & Distributed Processing (IPDPS), 2013 IEEE 27th International Symposium on. IEEE, 2013: 569-578.
8将RDMA应用于NFS的研究
引入RDMA提高性能,并解决了一个安全性问题。
Noronha R, Chai L, Talpey T, et al. Designing NFS with RDMA for security, performance and scalability[C]//Parallel Processing, 2007. ICPP 2007. International Conference on. IEEE, 2007: 49-49.
将RDMA作为NFS的传输层提高性能
Callaghan B, Lingutla-Raj T, Chiu A, et al. Nfs over rdma[C]//Proceedings of the ACM SIGCOMM workshop on Network-I/O convergence: experience, lessons, implications. ACM, 2003: 196-208.
分析发现基于RDMA的NFS在广域网上性能较差,由于广域网上RDMA读操作性能较差
Yu W, Rao N S V, Wyckoff P, et al. Performance of RDMA-capable storage protocols on wide-area network[C]//Petascale Data Storage Workshop, 2008. PDSW'08. 3rd. IEEE, 2008: 1-5.
9. 将RDMA应用于MPI的研究
Liu J, Wu J, Panda D K. High performance RDMA-based MPI implementation over InfiniBand[J]. International Journal of Parallel Programming, 2004, 32(3): 167-198.
Sur S, Jin H W, Chai L, et al. RDMA read based rendezvous protocol for MPI over InfiniBand: design alternatives and benefits[C]//Proceedings of the eleventh ACM SIGPLAN symposium on Principles and practice of parallel programming. ACM, 2006: 32-39.
- 10. 杂
还有一些研究RDMA编程,给web协议提速,在(IBMHPS (High Performance Switch and adapter)上实现RDMA等等