Distributed Storage and Consistency

1. Distributed Storage and Consistency

2. Storage moves into the net Network delays Network cost Storage capacity/volume Administrative cost Network bandwidth Shared storage with scalable bandwidth and capacity. Consolidate — multiplex — decentralize — replicate. Reconfigure to mix-and-match loads and resources.

3. Storage as a service SSP ASP Storage Service Provider Application Service Provider Outsourcing: storage and/or applications as a service. For ASPs (e.g., Web services), storage is just a component.

4. Storage Abstractions • relational database (IBM and Oracle) • tables, transactions, query language • file system • hierarchical name space of files with ACLs • Each file is a linear space of fixed-size blocks. • block storage • SAN, Petal, RAID-in-a-box (e.g., EMC) • Each logical unit (LU) or volume is a linear space of fixed-size blocks. • object storage • object == file, with a flat name space: NASD, DDS, Porcupine • Varying views of the object size: NASD/OSD/Slice objects may act as large-ish “buckets” that aggregate file system state. • persistent objects • pointer structures, requires transactions: OODB, ObjectStore

5. Network Block Storage • One approach to scalable storage is to attach raw block storage to a network. • Abstraction: OS addresses storage by <volume, sector>. • iSCSI, Petal, FibreChannel: access through special device driver • Dedicated Storage Area Network or general-purpose network. • FibreChannel (FC) vs. Ethernet • Volume-based administrative tools • backup, volume replication, remote sharing • Called “raw” or “block”, “storage volumes” or just “SAN”. • Least common denominator for any file system or database.

6. “NAS vs. SAN” • In the commercial sector there is a raging debate today about “NAS vs. SAN”. • Network-Attached Storage has been the dominant approach to shared storage since NFS. • NAS == NFS or CIFS: named files over Ethernet/Internet. • E.g., Network Appliance “filers” • Proponents of FibreChannel SANs market them as a fundamentally faster way to access shared storage. • no “indirection through a file server” (“SAD”) • lower overhead on clients • network is better/faster (if not cheaper) and dedicated/trusted • Brocade, HP, Emulex are some big players.

7. NAS vs. SAN: Cutting through the BS • FibreChannel a high-end technology incorporating NIC enhancements to reduce host overhead.... • ...but bogged down in interoperability problems. • Ethernet is getting faster faster than FibreChannel. • gigabit, 10-gigabit, + smarter NICs, + smarter/faster switches • Future battleground is Ethernet vs. Infiniband. • The choice of network is fundamentally orthogonal to storage service design. • Well, almost: flow control, RDMA, user-level access (DAFS/VI) • The fundamental questions are really about abstractions. • shared raw volume vs. shared file volume vs. private disks

8. Storage Architecture • Any of these abstractions can be built using any, some, or all of the others. • Use the “right” abstraction for your application. • Basic operations: create/remove, open/close, read/write. • The fundamental questions are: • What is the best way to build the abstraction you want? • division of function between device, network, server, and client • What level of the system should implement the features and properties you want?

9. Duke Mass Storage Testbed Goal: managed storage on demand for cross-disciplinary research. Direct SAN access for “power clients” and NAS PoPs; other clients access through NAS. IBM Shark/HSM Campus FC net IP LAN IP LAN Brain Lab Med Ctr

10. Problems • poor interoperability • Must have a common volume layout across heterogeneous SAN clients. • poor sharing control • The granularity of access control is an entire volume. • SAN clients must be trusted. • SAN clients must coordinate their access. • $$$

11. Campus IP net Duke Storage Testbed, v2.0 IBM Shark/HSM Each SAN volume is managed by a single NAS PoP. All access to each volume is mediated by its NAS PoP. Campus FC net Brain Lab Med Ctr

12. Testbed v2.0: pro and con • Supports resource sharing and data sharing. • Does not leverage Fibre Channel investment. • Does not scale access to individual volumes. • Prone to load imbalances. • Data crosses campus IP network in the clear. • Identities and authentication must be centrally administered. • It’s only as good as the NAS clients, which tend to be fair at best.

13. Sharing Network Storage • How can we control sharing to a space of files or blocks? • Access control etc. • Data model and storage abstraction • Caching • Optimistic replication • Consistency • One-copy consistency vs. weak consistency • Read-only (immutable) files? • Read-mostly files with weak consistency? • Write-anywhere files?

14. File/Block Cache Consistency • Basic write-ownership protocol. • Distributed shared memory (software DSM) • Timestamp validation (NFS). • Timestamp each cache entry, and periodically query the server: “has this file changed since time t?”; invalidate cache if stale. • Callback invalidation (AFS, Sprite, Spritely NFS). • Request notification (callback) from the server if the file changes; invalidate cache and/or disable caching on callback. • Leases (NQ-NFS, NFSv4, DAFS) • [Gray&Cheriton89,Macklem93]

15. switched interconnect Software DSM 101 • Software-based distributed shared memory (DSM) provides an illusion of shared memory on a cluster. • remote-fork the same program on each node • data resides in common virtual address space • library/kernel collude to make the shared VAS appear consistent • The Great War: shared memory vs. message passing • for the full story, take CPS 221

16. Page Based DSM (Shared Virtual Memory) • Virtual address space is shared Virtual Address Space physical physical DRAM DRAM

17. The Sequential Consistency Memory Model sequential processors issue memory ops in program order P3 P1 P2 switch randomly set after each memory op ensures some serial order among all operations Easily implemented with shared bus. For page-based DSM, weaker consistency models may be useful….but that’s for later. Memory

18. Inside Page-Based DSM (SVM) • The page-based approach uses a write-ownership token protocol on virtual memory pages. • Kai Li [Ivy SVM, 1986], Paul Leach [Apollo, 1982] • Each node maintains per-node per-page access mode. • {shared, exclusive, no-access} • determines local accesses allowed • For SVM, modes are enforced with VM page protection mode load (read) store (write) shared yes no exclusive yes yes no-access no no

19. Write-Ownership Protocol • A write-ownership protocol guarantees that nodes observe sequential consistency of memory accesses: • Any node with any access has the latest copy of the page. • On any transition from no-access, fetch current copy of page. • A node with exclusive access holds the only copy. • At most one node may hold a page in exclusive mode. • On transition into exclusive, invalidate all remote copies and set their mode to no-access. • Multiple nodes may hold a page in shared mode. • Permits concurrent reads: every holder has the same data. • On transition into shared mode, invalidate the exclusive remote copy (if any), and set its mode to shared as well.

20. Network File System (NFS) server client syscall layer user programs VFS syscall layer NFS server VFS *FS NFS client *FS RPC over UDP or TCP

21. NFS Protocol • NFS is a network protocol layered above TCP/IP. • Original implementations (and most today) use UDP datagram transport for low overhead. • Maximum IP datagram size was increased to match FS block size, to allow send/receive of entire file blocks. • Some implementations use TCP as a transport. • The NFS protocol is a set of message formats and types. • Client issues a request message for a service operation. • Server performs requested operation and returns a reply message with status and (perhaps) requested data.

22. File Handles • Question: how does the client tell the server which file or directory the operation applies to? • Similarly, how does the server return the result of a lookup? • More generally, how to pass a pointer or an object reference as an argument/result of an RPC call? • In NFS, the reference is a file handle or fhandle, a token/ticket whose value is determined by the server. • Includes all information needed to identify the file/object on the server, and find it quickly. volume ID inode # generation #

23. Consistency for File Systems • How is the consistency problem different for network file systems, relative to DSM/SVM? • Note: The CDK text includes a lot of detail about the kernel implementation issues for these file systems. These are interesting and useful, but in this course we focus on the distribution aspects.

24. NFS as a “Stateless” Service • A classical NFS server maintains no in-memory hard state. • The only hard state is the stable file system image on disk. • no record of clients or open files • no implicit arguments to requests • E.g., no server-maintained file offsets: read and write requests must explicitly transmit the byte offset for each operation. • no write-back caching on the server • no record of recently processed requests • etc., etc.... • “Statelessness makes failure recovery simple and efficient.”

25. Recovery in Stateless NFS • If the server fails and restarts, there is no need to rebuild in-memory state on the server. • Client reestablishes contact (e.g., TCP connection). • Client retransmits pending requests. • Classical NFS uses a connectionless transport (UDP). • Server failure is transparent to the client; no connection to break or reestablish. • A crashed server is indistinguishable from a slow server. • Sun/ONC RPC masks network errors by retransmitting a request after an adaptive timeout. • A dropped packet is indistinguishable from a crashed server.

26. Drawbacks of a Stateless Service • The stateless nature of classical NFS has compelling design advantages (simplicity), but also some key drawbacks: • Recovery-by-retransmission constrains the server interface. • ONC RPC/UDP has execute-mostly-once semantics (“send and pray”), which compromises performance and correctness. • Update operations are disk-limited. • Updates must commit synchronously at the server. • NFS cannot (quite) preserve local single-copy semantics. • Files may be removed while they are open on the client. • Server cannot help in client cache consistency. • Let’s look at the consistency problem...

27. Timestamp Validation in NFS [1985] • NFSv2 and NFSv3 cache consistency uses a form of timestamp validation like today’s Web • Timestamp cached data at file grain. • Maintain per-file expiration time (TTL) • Probe for new timestamp to revalidate if cache TTL has expired. • Get attributes (getattr) • Key difference: NFS file cache and access primitives are block-grained, and the client may issue many operations in sequence on the same file. • Clustering: File-grained timestamp for block-grained cache • Piggyback file attributes on each response • Adaptive TTL • What happens on server failure? Client failure?

28. AFS [1985] • AFS is an alternative to NFS developed at CMU. • Duke still uses it. • Designed for wide area file sharing: • Internet is large and growing exponentially. • Global name hierarchy with local naming contexts and location info embedded in fully qualified names. • Much like DNS • Security features, with per-domain authentication / access control. • Whole file caching or 64KB chunk caching • Amortize request/transfer cost • Client uses a disk cache • Cache is preserved across client failure. • Again, it looks a lot like the Web.

29. Callback Invalidations in AFS-2 • AFS-1 uses timestamp validation like NFS; AFS-2 uses callback invalidations. • Server returns “callback promise” token with file access. • Like ownership protocol, confers a right to cache the file. • Client caches the token on its disk. • Token states: {valid, invalid, cancelled} • On a sharing collision, server cancels token with a callback. • Client invalidates cached copy of the associated file. • Detected on client write to server: last writer wins. • (No distinction between read/write token.)

30. Issues with AFS Callback Invalidations • What happens after a failure? • Client invalidates its tokens on client restart. • Invalid tokens may be revalidated, like NFS getattr or WWW. • Server must remember tokens across server restart. • Can the client distinguish a server failure from a network failure? • Client invalidates tokens after a timeout interval T if the client has no communication with the server. • Weakens consistency in failures. • Then there’s the problem of update semantics: two clients may be actively updating the same file at the same time.

31. NQ-NFS Leases • In NQ-NFS, a client obtains a lease on the file that permits the client’s desired read/write activity. • “A lease is a ticket permitting an activity; the lease is valid until some expiration time.” • A read-caching lease allows the client to cache clean data. • Guarantee: no other client is modifying the file. • A write-caching lease allows the client to buffer modified data for the file. • Guarantee: no other client has the file cached. • Allows delayed writes: client may delay issuing writes to improve write performance (i.e., client has a writeback cache).

32. Using NQ-NFS Leases • 1. Client NFS piggybacks lease requests for a given file on I/O operation requests (e.g., read/write). • NQ-NFS leases are implicit and distinct from file locking. • 2. The server determines if it can safely grant the request, i.e., does it conflict with a lease held by another client. • read leases may be granted simultaneously to multiple clients • write leases are granted exclusively to a single client • 3. If a conflict exists, the server may send an eviction notice to the holder of the conflicting lease. • If a client is evicted from a write lease, it must write back. • Grace period: server grants extensions while the client writes. • Client sends vacated notice when all writes are complete.

33. NQ-NFS Lease Recovery • Key point: the bounded lease term simplifies recovery. • Before a lease expires, the client must renew the lease. • What if a client fails while holding a lease? • Server waits until the lease expires, then unilaterally reclaims the lease; client forgets all about it. • If a client fails while writing on an eviction, server waits for write slack time before granting conflicting lease. • What if the server fails while there are outstanding leases? • Wait for lease period + clock skew before issuing new leases. • Recovering server must absorb lease renewal requests and/or writes for vacated leases.

34. NQ-NFS Leases and Cache Consistency • Every lease contains a file version number. • Invalidation cache iff version number has changed. • Clients may disable client caching when there is concurrent write sharing. • no-caching lease (Sprite) • What consistency guarantees do NQ-NFS leases provide? • Does the server eventually receive/accept all writes? • Does the server accept the writes in order? • Are groups of related writes atomic? • How are write errors reported? • What is the relationship to NFS V3 commit?

35. The Distributed Lock Lab • The lock implementation is similar to DSM systems, with reliability features similar to distributed file caches. • use Java RMI • lock token caching with callbacks • lock tokens passed through server, not peer-peer as DSM • synchronizes multiple threads on same client • state bit for pending callback on client • server must reissue callback each lease interval (or use RMI timeouts to detect a failed client) • client must renew token each lease interval

36. Remote Method Invocation (RMI) RMI is “RPC in Java”, supporting Emerald-like distributed object references, invocation, and garbage collection, derived from SRC Modula-3 network objects [SOSP 93]. RMI registry The registry provides a bootstrap naming service using URLs. obj1 obj2 rmi://slowww.server.edu/object1 obj3 1: Naming.bind(URL, obj1) 2: stub1 = Naming.lookup(URL) server app client app 3: stub2 = stub1->method() skeleton stub RMI layer RMI layer transport transport server VM client VM

37. Background Slides These slides were not discussed. I use them in CPS 210, the operating systems course. They provide useful background for the material on NFS.

38. cluster FS cluster FS Cluster File Systems storage client storage client shared block storage service (FC/SAN, Petal, NASD) xFS [Dahlin95] Petal/Frangipani [Lee/Thekkath] GFS Veritas EMC Celerra issues trust compatibility with NAS protocols sharing, coordination, recovery

39. *FS client *FS client *FS svc *FS svc Sharing and Coordination block allocation and layout locking/leases, granularity shared access separate lock service logging and recovery network partitions reconfiguration NAS “SAN” storage service + lock manager What does Frangipani need from Petal? How does Petal contribute to F’s *ility? Could we build Frangipani without Petal?

40. A Typical Unix File Tree Each volume is a set of directories and files; a host’s file tree is the set of directories and files visible to processes on a given host. / File trees are built by grafting volumes from different volumes or from network servers. bin etc tmp usr vmunix In Unix, the graft operation is the privileged mount system call, and each volume is a filesystem. ls sh project users packages mount point • mount (coveredDir, volume) • coveredDir: directory pathname • volume: device specifier or network volume • volume root contents become visible at pathname coveredDir (volume root) tex emacs

41. Filesystems • Each file volume (filesystem) has a type, determined by its disk layout or the network protocol used to access it. • ufs (ffs), lfs, nfs, rfs, cdfs, etc. • Filesystems are administered independently. • Modern systems also include “logical” pseudo-filesystems in the naming tree, accessible through the file syscalls. • procfs: the /proc filesystem allows access to process internals. • mfs: the memory file system is a memory-based scratch store. • Processes access filesystems through common system calls.

42. user space syscall layer (file, uio, etc.) Virtual File System (VFS) network protocol stack (TCP/IP) FFS LFS NFS *FS etc. etc. device drivers VFS: the Filesystem Switch • Sun Microsystems introduced the virtual file system interface in 1985 to accommodate diverse filesystem types cleanly. • VFS allows diverse specific file systems to coexist in a file tree, isolating all FS-dependencies in pluggable filesystem modules. VFS was an internal kernel restructuring with no effect on the syscall interface. Incorporates object-oriented concepts: a generic procedural interface with multiple implementations. Based on abstract objects with dynamic method binding by type...in C. Other abstract interfaces in the kernel: device drivers, file objects, executable files, memory objects.

43. syscall layer free vnodes NFS UFS Vnodes • In the VFS framework, every file or directory in active use is represented by a vnode object in kernel memory. Each vnode has a standard file attributes struct. Generic vnode points at filesystem-specific struct (e.g., inode, rnode), seen only by the filesystem. Each specific file system maintains a cache of its resident vnodes. Vnode operations are macros that vector to filesystem-specific procedures.

44. Vnode Operations and Attributes vnode attributes (vattr) type (VREG, VDIR, VLNK, etc.) mode (9+ bits of permissions) nlink (hard link count) owner user ID owner group ID filesystem ID unique file ID file size (bytes and blocks) access time modify time generation number directories only vop_lookup (OUT vpp, name) vop_create (OUT vpp, name, vattr) vop_remove (vp, name) vop_link (vp, name) vop_rename (vp, name, tdvp, tvp, name) vop_mkdir (OUT vpp, name, vattr) vop_rmdir (vp, name) vop_symlink (OUT vpp, name, vattr, contents) vop_readdir (uio, cookie) vop_readlink (uio) files only vop_getpages (page**, count, offset) vop_putpages (page**, count, sync, offset) vop_fsync () generic operations vop_getattr (vattr) vop_setattr (vattr) vhold() vholdrele()

45. V/Inode Cache VFS free list head HASH(fsid, fileid) Active vnodes are reference- counted by the structures that hold pointers to them. - system open file table - process current directory - file system mount points - etc. Each specific file system maintains its own hash of vnodes (BSD). - specific FS handles initialization - free list is maintained by VFS vget(vp): reclaim cached inactive vnode from VFS free list vref(vp): increment reference count on an active vnode vrele(vp): release reference count on a vnode vgone(vp): vnode is no longer valid (file is removed)

46. Pathname Traversal • When a pathname is passed as an argument to a system call, the syscall layer must “convert it to a vnode”. • Pathname traversal is a sequence of vop_lookup calls to descend the tree to the named file or directory. Issues: 1. crossing mount points 2. obtaining root vnode (or current dir) 3. finding resident vnodes in memory 4. caching name->vnode translations 5. symbolic (soft) links 6. disk implementation of directories 7. locking/referencing to handle races with name create and delete operations open(“/tmp/zot”) vp = get vnode for / (rootdir) vp->vop_lookup(&cvp, “tmp”); vp = cvp; vp->vop_lookup(&cvp, “zot”);

47. Problem 1: Retransmissions and Idempotency • For a connectionless RPC transport, retransmissions can saturate an overloaded server. • Clients “kick ‘em while they’re down”, causing steep hockey stick. • Execute-at-least-once constrains the server interface. • Service operations should/must be idempotent. • Multiple executions should/must have the same effect. • Idempotent operations cannot capture the full semantics we expect from our file system. • remove, append-mode writes, exclusive create

48. Solutions to the Retransmission Problem • 1. Hope for the best and smooth over non-idempotent requests. • E.g., map ENOENT and EEXIST to ESUCCESS. • 2. Use TCP or some other transport protocol that produces reliable, in-order delivery. • higher overhead...and we still need sessions. • 3. Implement an execute-at-most once RPC transport. • TCP-like features (sequence numbers)...and sessions. • 4. Keep a retransmission cache on the server [Juszczak90]. • Remember the most recent request IDs and their results, and just resend the result....does this violate statelessness? • DAFS persistent session cache.

49. Problem 2: Synchronous Writes • Stateless NFS servers must commit each operation to stable storage before responding to the client. • Interferes with FS optimizations, e.g., clustering, LFS, and disk write ordering (seek scheduling). • Damages bandwidth and scalability. • Imposes disk access latency for each request. • Not so bad for a logged write; much worse for a complex operation like an FFS file write. • The synchronous update problem occurs for any storage service with reliable update (commit).

50. Speeding Up Synchronous NFS Writes • Interesting solutions to the synchronous write problem, used in high-performance NFS servers: • Delay the response until convenient for the server. • E.g., NFS write-gathering optimizations for clustered writes (similar to group commit in databases). • Relies on write-behind from NFS I/O daemons (iods). • Throw hardware at it: non-volatile memory (NVRAM) • Battery-backed RAM or UPS (uninterruptible power supply). • Use as an operation log (Network Appliance WAFL)... • ...or as a non-volatile disk write buffer (Legato). • Replicate server and buffer in memory (e.g., MIT Harp).