Distributed Systems CS 15 440 Distributed File Systems

Distributed Systems CS 15 -440 Distributed File Systems- Part I Lecture 20, Nov 28, 2012 Majd F. Sakr and Mohammad Hammoud 1

Today… § Last 3 sessions § Fault Tolerance § Today’s session § Distributed File Systems – Part I § Announcements: § Project 2 grades are out § PS 4 (the last assignment) is now posted. Due date: Dec 12, 2012 § Project 4 (the last project) will be posted by tonight 2

Intended Learning Outcomes: Distributed File Systems Considered: a reasonably critical and comprehensive perspective. Thoughtful: Fluent, flexible and efficient perspective. Masterful: a powerful and illuminating perspective. Explain distributed file systems as a paradigm for general-purpose distributed systems, and analyze its various aspects and architectures ILO 7. 1 Define distributed file systems (DFSs) and explain various architectures ILO 7. 2 Analyze various aspects of DFSs including ILO 7. 1 ILO 7. 2 processes, communication, naming, synchronization, consistency and replication, and fault tolerance

Discussion on Distributed File Systems (DFSs) Basics DFS Aspects 4

Distributed File Systems § Why File Systems? § To organize data (as files) § To provide a means for applications to store, access, and modify data § Why Distributed File Systems (DFSs)? § Sharing and managing data in distributed systems § Big data continues to grow § A DFS can hold sheer volumes of data (contrary to a local file system) and provide access to these data to many clients dispersed across networks

NAS versus SAN § Another term for DFS is network attached storage (NAS), referring to attaching storage to network servers that provide file systems § A similar sounding term that refers to a very different approach is storage area network (SAN) § SAN makes storage devices (not file systems) available over a network Client Computer File Server (Providing NAS) Client Computer LAN SAN Database Server

Benefits of DFSs § DFSs provide: 1. File sharing over a network: without a DFS, we would have to exchange files by e-mail or use applications such as the Internet’s FTP 2. Transparent files accesses: A user’s programs can access remote files as if they are local. The remote files have no special APIs; they are accessed just like local ones 3. Easy file management: managing a DFS is easier than managing multiple local file systems

DFS Components § DFS information can be typically categorized as follows: 1. The data state: This is the contents of files 2. The attribute state: This is the information about each file (e. g. , file’s size and access control list) 3. The open-file state: This includes which files are open or otherwise in use, as well as describing how files are locked § Designing a DFS entails determining how its various components are placed. Specifically, by component placement we indicate: § § What resides on the servers What resides on the clients

DFS Component Placement (1) § The data and the attribute states permanently reside on the server’s local file system, but recently accessed or modified information might reside in server and/or client caches § The open-file state is transitory; it changes as processes open and close files Data Cache Attribute Cache Open-File State Client Data Cache Network Attribute Cache Local File System Open-File State Server

DFS Component Placement (2) § Three basic concerns placement strategy: govern the DFS components 1. Access speed: Caching information on clients improves performance considerably 2. Consistency: If clients cache information, do all parties share the same view of it? 3. Recovery: If one or more computers crash, to what extent are the others affected? How much information is lost?

Discussion on Distributed File Systems (DFSs) Basics DFS Aspects 11

DFS Aspects Aspect Description Architecture How are DFSs generally organized? Processes • • Who are the cooperating processes? Are processes stateful or stateless? Communication • What is the typical communication paradigm followed by DFSs? How do processes in DFSs communicate? • Naming How is naming often handled in DFSs? Synchronization What are the file sharing semantics adopted by DFSs? Consistency and Replication What are the various features of client-side caching as well as server-side replication? Fault Tolerance How is fault tolerance handled in DFSs?

Architectures § Client-Server Distributed File Systems § Cluster-Based Distributed File Systems § Symmetric Distributed File Systems

Network File System § Many distributed file systems are organized along the lines of client-server architectures § Sun Microsystem’s Network File System (NFS) is one of the most widely-deployed DFS for Unix-based systems § NFS comes with a protocol that describes precisely how a client can access a file stored on a (remote) NFS file server § NFS allows a heterogeneous collection of processes, possibly running on different OSs and machines, to share a common file system

Remote Access Model § The model underlying NFS and similar systems is that of remote access model Replies from server Server Client File Requests from client to access remote file File stays on server § In this model, clients: § Are offered transparent access to a file system that is managed by a remote server § Are normally unaware of the actual location of files § Are offered an interface to a file system similar to the interface offered by a conventional local file system

Upload/Download Model § A contrary model, referred to as upload/download model, allows a client to access a file locally after having downloaded it from the server File moved to client Server Client File New File All accesses are done on client When client is done, file is returned to server § The Internet’s FTP service can be used this way when a client downloads a complete file, modifies it, and then puts it back

The Basic NFS Architecture Client Server System call layer Virtual File System (VFS) layer Local file system interface NFS client NFS server RPC client stub RPC server stub A Client Request in NFS Network Local file system interface

Architectures § Client-Server Distributed File Systems § Cluster-Based Distributed File Systems § Symmetric Distributed File Systems

Data-Intensive Applications § Today there is a deluge of large data-intensive applications § Most data-intensive applications fall into one of two styles of computing: § Internet services (or cloud computing) § High-performance computing (HPC) § Cloud computing and HPC applications: § Run typically on thousands of compute nodes Visualization of entropy in Terascale § Process sheer volumes of data Supernova Initiative application. Image (or big data) from Kwan-Liu Ma’s visualization team at UC Davis

Cluster-Based Distributed File Systems § The underlying cluster-based file system is a key component for providing scalable data-intensive application performance § The cluster-based file system divides and distributes big data, using file striping techniques, for allowing concurrent data accesses § The cluster-based file systems could be either cloud computing or HPC oriented § Examples: § Cloud Computing Oriented: Google File System (GFS) § HPC Oriented: Parallel Virtual File System (PVFS)

File Striping Techniques § Server clusters are often used for distributed applications and their associated file systems are adjusted to satisfy their requirements § One well-known technique is to deploy file-striping techniques, by which a single file is distributed across multiple servers § Hence, it becomes possible to fetch different parts concurrently Accessing file parts in parallel a b c e d a c e a b b d d e c

Round-Robin Distribution (1) § How to stripe a file over multiple machines? § Round-Robin is typically a reasonable default solution Logical File 0 1 2 3 4 5 6 7 1 5 9 13 8 9 10 11 12 13 14 15 Stripe Size Striping Unit 0 4 8 Server 1 12 Server 2 2 6 10 Server 3 14 3 7 11 Server 4 15

Round-Robin Distribution (2) § Clients perform writes/reads of file at various regions Client I: 512 K write, offset 0 0 0 1 4 0 2 1 4 8 Server 1 3 Client II: 512 K write, offset 512 4 5 6 5 2 6 12 1 5 7 8 3 9 Server 2 7 13 8 2 9 12 6 10 9 10 Server 3 14 11 12 13 13 10 3 14 14 7 11 Server 4 15 11 15 15

2 D Round-Robin Distribution (1) § What happens when we have many servers (say 1000 s)? § 2 D distribution can help Logical File 0 1 2 3 4 5 6 7 1 3 5 7 8 9 10 11 12 13 14 15 Stripe Size Striping Unit 0 2 4 Server 1 Group Size = 2 6 Server 2 8 10 12 Server 3 14 9 11 13 Server 4 15

2 D Round-Robin Distribution (2) § 2 D distribution can limit the number of servers per client Client I: 512 K write, offset 0 0 0 1 4 0 2 2 3 Client II: 512 K write, offset 512 4 5 6 1 5 2 6 4 6 1 3 Server 1 Group Size = 2 7 8 3 5 Server 2 7 7 8 8 9 12 10 10 9 12 Server 3 14 11 12 13 13 10 9 11 14 14 13 Server 4 15 11 15 15

GFS Data Distribution Policy § The Google File System (GFS) is a scalable DFS for data-intensive applications § GFS divides large files into multiple pieces called chunks or blocks (by default 64 MB) and stores them on different data servers § This design is referred to as block-based design § Each GFS chunk has a unique 64 -bit identifier and is stored as a file in the lower-layer local file system on the data server § GFS distributes chunks across cluster data servers using a random distribution policy

GFS Random Distribution Policy Server 0 (Writer) Server 1 Server 2 Server 3 0 M Blk 0 Blk 1 Blk 0 64 M Blk 1 Blk 2 Blk 1 128 M Blk 2 Blk 3 Blk 4 192 M Blk 3 Blk 6 256 M Blk 4 Blk 5 320 M Blk 5 384 M Blk 6 Load Imbalance

GFS Architecture § The storage and compute capabilities of a cluster are organized in two ways: 1. 2. GFS Co-locate storage and compute in the same node Separate storage nodes from compute nodes File name, chunk index GFS client Master Contact address Chunk Id, range Chunk data Chunk Server Linux File System

PVFS Data Distribution Policy § Parallel Virtual File System (PVFS) is a scalable DFS for (scientific) data-intensive applications § PVFS divides large files into multiple pieces called stripe units (by default 64 KB) and stores them on different data servers § This design is referred to as object-based design § Unlike the block-based design of GFS, PVFS stores an object (or a handle) as a file that includes all the stripe units at a data server § PVFS distributes stripe units across cluster data servers using a round robin policy

PVFS Round-Robin Distribution Policy Server 0 (Writer) Server 1 Server 2 Server 3 0 M 64 M Blk 0 Blk 1 128 M Blk 1 Blk 2 192 M Blk 2 Blk 3 Blk 4 Blk 5 Blk 6 256 M 320 M 384 M Blk 6 Load Balance

PVFS Architecture § The storage and compute capabilities of a cluster are organized in two ways: 1. 2. Co-locate storage and compute in the same node Separate storage nodes. PVFS from compute nodes I/O Nodes Network Compute Nodes Metadata Manager

Architectures § Client-Server Distributed File Systems § Cluster-Based Distributed File Systems § Symmetric Distributed File Systems

Ivy § Fully symmetric organizations that are based on peer-to-peer technology also exist § All current proposals use a DHT-based system for distributing data, combined with a key-based lookup mechanism § As an example, Ivy is a distributed file system that is built using a Chord DHT-based system § Data storage in Ivy is realized by a block-oriented (i. e. , blocks are distributed over storage nodes) distributed storage called DHash

Ivy Architecture § Ivy consists of 3 separate layers: Node where a file system is rooted File System Layer Block-Oriented Storage DHT Layer Ivy Ivy DHash Chord

Ivy § Ivy implements an NFS-like file system § To increase availability and improve performance, Ivy: § Replicates every block B to the k immediate successors of the server responsible for storing B § Caches looked up blocks along the route that the lookup request followed

DFS Aspects Aspect Description Architecture How are DFSs generally organized? Processes • • Who are the cooperating processes? Are processes stateful or stateless? Communication • What is the typical communication paradigm followed by DFSs? How do processes in DFSs communicate? • Naming How is naming often handled in DFSs? Synchronization What are the file sharing semantics adopted by DFSs? Consistency and Replication What are the various features of client-side caching as well as server-side replication? Fault Tolerance How is fault tolerance handled in DFSs?

Stateless Processes § Cooperating Processes in DFSs are usually the storage servers and file manager(s) § The most important aspect concerning DFS processes is whether they should be stateless or stateful § The Stateless Approach: § Doesn’t require that servers maintain any client state § When a server crashes, there is essentially no need to enter a recovery phase to bring the server to a previous state § Locking a file cannot be easily done § E. g. , NFSv 3

Stateful Processes § The Stateful Approach: § Requires that a server maintains some client state § Clients can make effective use of caches but this would require a cache consistency protocol § Provides a server with the ability to support callbacks (i. e. , the ability to do RPC to a client) in order to keep track of its clients § E. g. , NFSv 4

DFS Aspects Aspect Description Architecture How are DFSs generally organized? Processes • • Who are the cooperating processes? Are processes stateful or stateless? Communication • What is the typical communication paradigm followed by DFSs? How do processes in DFSs communicate? • Naming How is naming often handled in DFSs? Synchronization What are the file sharing semantics adopted by DFSs? Consistency and Replication What are the various features of client-side caching as well as server-side replication? Fault Tolerance How is fault tolerance handled in DFSs?

Communication in DFSs § Communication in DFSs is mainly based on remote procedure calls (RPCs) § The main reason for choosing RPC is to make the system independent from underlying OSs, networks, and transport protocols § GFS uses RPC and may break a read into multiple RPCs to increase parallelism § PVFS currently uses TCP for all its internal communication § In NFS, all communication between a client and server proceeds along the Open Network Computing RPC (ONC RPC)

RPCs in NFS Client Lookup § Up until NFSv 4, the client was made responsible for making the server’s life as easy as possible by keeping requests simple § In that case, the extra latency of a second RPC leads to performance degradation Lookup name Time § The drawback becomes apparent when considering the use of NFS in a wide-area system Server Read file data Client Server Lookup Open Read § To circumvent such a problem, NFSv 4 supports compound procedures Time Lookup name Open file Read file data

RPC in Coda: RPC 2 § Another enhancement to RPCs has been developed as part of the Coda file system (Kistler and Satyanarayanan, 1992) and referred to as RPC 2 § RPC 2 is a package that offers reliable RPCs on top of the (unreliable) UDP protocol § Each time an RPC is made, the RPC 2 client code starts a new thread that sends an invocation request to the server then blocks until it receives an answer § As request processing may take an arbitrary time to complete, the server regularly sends back messages to the client to let it know it is still working on the request

RPC 2 and Multicasting § RPC 2 supports multicasting § An important design issue in Coda is that servers keep track of which clients have a local copy of a file (i. e. , stateful servers) § When a file is modified, a sever invalidates local copies at clients in parallel using multicasting Client Invalidate Reply Client Reply Server Invalidate Reply Client Time Sending an invalidation message one at a time Reply Time Sending an invalidation message using multicasting as in RPC 2

Next Class: DFS Aspects Aspect Description Architecture How are DFSs generally organized? Processes • • Who are the cooperating processes? Are processes stateful or stateless? Communication • What is the typical communication paradigm followed by DFSs? How do processes in DFSs communicate? • Naming How is naming often handled in DFSs? Synchronization What are the file sharing semantics adopted by DFSs? Consistency and Replication What are the various features of client-side caching as well as server-side replication? Fault Tolerance How is fault tolerance handled in DFSs?