Chapter 17 Distributed Communication

Distributed File Systems 17 CHAPTER Exercises 17.1 What are the benefits of a DFS when compared to a file system in a centralized system? Answer: A DFS allows the same type of sharing available on a centralized system, but the sharing may occur on physically and logically separate systems. Users across the world are able to share data as if they were in the same building, allowing a much more flexible computing environment than would otherwise be available. 17.2 Which of the example DFSs discussed in this chapter would handle a large, multiclient database application most efficiently? Explain your answer. Answer: The Andrew file system can handle a large, multiclient database as scalability is one of its hallmark features. Andrew is designed to handle up to 5,000 client workstations as well. A database also needs to run in a secure environment and Andrew uses the Kerberos security mechanism for encryption. 17.3 Discuss whether AFS and NFS provide the following: (a) location transparency and (b) location independence. Answer: NFS provides location transparence since one cannot determine the server hosting the file from its name. (One could however view the mount-tables to determine the file server from which the corresponding file system is mounted, but the file server is not hardcoded in the name of the file.) NFS does not provide location independence since files can not be moved automatically between different file systems. AFS provides location transparence and location independence. 117 118 Chapter 17 Distributed File Systems 17.4 Under what circumstances would a client prefer a location-transparent DFS? Under which circumstances would she prefer a location-independent DFS? Discuss the reasons for these preferences. Answer: Location-transparent DFS is good enough in systems in which files are not replicated. Location-independent DFS is necessary when any replication is done. 17.5 What aspects of a distributed system would you select for a system running on a totally reliable network? Answer: Since the system is totally reliable, a stateful approach would make the most sense. Error recovery would seldom be needed, allowing the features of a stateful system to be used. If the network is very fast as well as reliable, caching can be done on the server side. On a slower network caching on both server and client will speed performance, as would file location-independence and migration. In addition, RPCbased service is not needed in the absence of failures, since a key part of its design is recovery during networking errors. Virtual-circuit systems are simpler and more appropriate for systems with no communications failures. 17.6 Consider AFS, which is a stateful distributed file system. What actions need to be performed to recover from a server crash in order to preserve the consistency guaranteed by the system? Answer: A server needs to keep track of what clients are currently caching a file in order to issue a callback when the file is modified. When a server goes down, this state is lost. A server would then have to reconstruct this state typically by contacting all of the clients and having them report to the server what files are currently being cached by each client. 17.7 Compare and contrast the techniques of caching disk blocks locally, on a client system, and remotely, on a server. Answer: Caching locally can reduce network traffic substantially as the local cache can possibly handle a significant number of the remote accesses. This can reduce the amount of network traffic and lessen the load on the server. However, to maintain consistency, local updates to disk blocks must be updated on the server using either a write-through or delayed-write policy. A strategy must also be provided that allows the client to determine if its cached data is stale and needs to be updated. Caching locally provides is obviously more complicated than having a client request all data from the server. But if access patterns indicate heavy writes to the data, the mechanisms for handling inconsistent data may increase network traffic and server load. 17.8 AFS is designed to support a large number of clients. Discuss three techniques used to make AFS a scalable system. Answer: Three techniques that make AFS a scalable system are: • Caching: AFS performs caching of files and name translations thereby limiting the number of operations sent to the server. • Whole-file caching: when a file is opened, the entire contents of the file is transported to the client and no further interactions with the Exercises 119 server is required. (This approach is refined in later versions where large chunks of a file rather than the entire file is transported in a single operation.) • Callbacks: It is the server’s responsibility to revoke outdated copies of files. Clients can cache files and reuse the cached data multiple times without making requests to the server. 17.9 Discuss the advantages and disadvantages of performing path-name translation by having the client ship the entire path to the server requesting a translation for the entire path name of the file. Answer: The advantage is that a single network request is sufficient to perform the path-name translation. Schemes that perform translations one component at a time incur more network traffic. However, when translations are performed one component at a time, the translations of the parent directories are obtained and cached for future reuse, whereas if the translation of the entire path is performed, translations for none of the intermediate elements are available in the cache. 17.10 What are the benefits of mapping objects into virtual memory, as Apollo Domain does? What are the drawbacks? Answer: Mapping objects into virtual memory greatly eases the sharing of data between processes. Rather than opening a file, locking access to it, and reading and writing sections via the I/O system calls, memorymapped objects are accessible as “normal” memory, with reads and writes to locations independent of disk pointers. Locking is much easier also, since one shared memory location can be used as a locking variable for semaphore access. Unfortunately, memory mapping adds complexity to the operating system, especially in a distributed system. 17.11 Describe some of the fundamental differences between AFS and NFS (see Chapter 11). Answer: Some of the distinctive differences include: • AFS has a rich set of features whereas NFS is organized around a much simpler design. • NFS allows a workstation to act as either a client, a server, or both. AFS distinguishes between clients and server and identifies dedicated servers. • NFS is stateless meaning a server does not maintain state during client updates of a file. AFS is stateful between the period of when a client opens a file, updates it, and closes the file. (NFS does not even allow the opening and closing of files.) • Caching is a fundamental feature of AFS allowing client-side caching with cache consistency. In fact, it is an architectural principle behind AFS to allow clients to cache entire files. Consistency is provided by servers when cached files are closed. The server then invalidates cached copies existing on other clients. Caching is allowed in NFS as well, but because of its stateless nature modified data must be committed to the server before results are received back by the client. 120 Chapter 17 Distributed File Systems • AFS provides session semantics whereas NFS supports UNIX file consistency semantics. 17.12 Discuss whether clients in the following systems can obtain inconsistent or stale data from the file server and, if so, under what scenarios this could occur. a. AFS b. Sprite c. NFS Answer: Sprite guarantees that clients never see stale data. When a file is opened for write-sharing by multiple clients, all caching for the file is disabled and all operations are sent directly to the server. This scheme guarantees consistency of data. In AFS, writes to files are performed on local disks, and when the client closes the files, the writes are propagated to the server, which then issues callbacks on the various cached copies. During the time the file is written but not closed, the other clients could be accessing stale data. Also even after the file is closed, the other clients might access stale data if they had performed the open on the file before the updating client had closed it. It is only at the point of time when the next open is performed on a caching client that the server is contacted and the most recent data propagated from the server to the client. NFS uses a more ad-hoc consistency mechanism. Data is flushed from the clients to servers at periodic intervals and on file close operations. Also a client caching the file data checks for inconsistencies also at periodic intervals. Any updates made and flushed to the server during these intervals are not seen immediately on a client caching the file data.