NFS in Userspace: Goals and Challenges

Transcription

1 NFS in Userspace: Goals and Challenges Tai Horgan EMC Isilon Storage Division 2013 Storage Developer Conference. Insert Your Company Name. All Rights Reserved.

2 Introduction: OneFS Clustered NAS File Server Single Volume Multiple Protocol Access SMB 1/2/2.1 NFS 2/3/4 HTTP FTP Hadoop Hybrid NTFS/Posix semantics

3 NFS in Userspace Why userspace? 1. Single Point of Access 2. Unified Lock Domains 3. Data Protection 4. Lower Maintenance Cost *

4 NFS in Userspace: Why? Single Point of Entry to Filesystem Protocol Drivers (SMB) SMB, NFS, FTP, etc Unified access auditing for all protocols Prevent unnecessary duplication of features Maintainable API into filesystem IoManager Audit Filter Driver OneFS Filesystem

5 NFS in Userspace: Why? Unified Lock Domain SMB already in userspace Correctly contend with NFS4 Best effort contention with NLM NFS4 Mandatory Locking SMB Mandatory Locking NLM Advisory Locking IoManager

6 NFS in Userspace: Why? OneFS Clustered Filesystem Reboot implies potential loss of data protection Reboots are more than panics - userland is easily patched

7 IoManager IoManager: Virtualized I/O Framework Windows-style, open-based file semantics Developed by Likewise Software, now a part of EMC/Isilon

8 IoManager Past Present Future SMB NFS3 NFS4.1 SMB2 NFS4.2 NFS4 SMB2.1 FTP/HTTP IoManager

9 NFS in Userspace Challenges: Performance Lock Semantics IoManager Idiosyncrasies NFS Filehandle Compatibility

10 Performance Usermode is Slow! Three main mitigation strategies: Aggressive Caching Custom RPC with Zero-Copy Microthreading *

11 Performance: Caching Client Cache invalidation: NFSv3: WCC data NFSv4: Delegations SMB: Oplocks/Leases Multiprotocol Stack: Simulate WCC with delegation/oplock on behalf of client *

12 Performance: Caching Client NFS Attribute Cache Client Write Pre-WCC Acquire Oplock Charge From Disk WRITE op Perform Write Post-WCC Update Size and Time Attributes OneFS Client Stat GETATTR op *

13 Performance: RPC Custom User-Mode RPC code Nonblocking Zero-Copy aware Custom.x file with read and write buffers demarcated Custom fields are replaced with socket pair in place of buffer Fd pairs are passed in/out of kernel

14 Performance: Microthreading IoManager lightweight thread model IoManager Fiber API 1-1 Threads to Cores ratio Nonblocking behind the scenes - blocking calls are fiber context switches and callbacks

15 Lock Semantics IoManager Mandatory share mode locks SMB-style control path: ops begin with OPEN/CREATE NFS Advisory locking Majority of ops are simple stats, which cannot fail for access violations

16 Lock Semantics Solution: More caching Cache opens for recent files Cached opens contend on share mode locks Non-conflicting fallback for operations that cannot return access errors (GETATTR)

17 Lock Semantics GETATTR Open Cache Existing Open WCC Cache New Open Attempt New Open OneFS Nonconflicting Open

18 IoManager Special Cases NFS REMOVE and delete-on-close Immediate stat should fail REMOVE of link should delete link, not target Non-regular file types Posix filenames

19 Filehandle Compatibility NFS Model Filehandle - unique file ID OneFS: Inode # / Snapshot ID / Export ID / FSID IoManager Path based API Full path Parent open + relative path

20 Filehandle Compatibility Hybrid model - Multiple combinations Filehandle only GETATTR, READ, WRITE Filehandle + relative path LOOKUP, NFSv4 Full path SMB, MOUNT Path + relative path SMB NFSv3 forced to suffer an extra OPEN for parent directory - offset by caching

21 NFS4-Specific Challenges Generally much easier than v3 Open semantics for READ and WRITE Mandatory locking NFS4 CurFH state provides natural parent open for relative paths

22 NFS4-Specific Challenges Atomicity of open upgrade and downgrade v4 open semantics require unioning of access on multiple opens of a single file Dropping and re-acquiring is racey Determining if we even have an open is difficult, due to hard links IoManager No upgrade or downgrade, but provides a GUID key to atomically replace an open

23 NFS4-Specific Challenges OPEN Follow Link New Open Open List Existing Open Compare Access Reopen with matching GUID New Open

24 NFS4-Specific Challenges Cluster Node Failover OneFS supports arbitrarily moving IP addresses from one cluster node to another IP Migration should be transparent to the client - service must not be affected Similar to NFSv4 Migration, though without the use of fs_location attribute

25 NFS4-Specific Challenges Migration of NFSv4 State Cannot halt cluster locking to allow lock reclaim Client state must be transferred with address Synchronize servers on per-ip basis Immigrating and emigrating servers must synchronize state before migration occurs No such scheme exists in OneFS yet

26 NFS in Userspace Goals: Performance parity Flesh out NFSv4 implementation (current server lacks delegation support) Where we stand now: Performance tuning Ironing out edge cases Plenty of work to do!

27 QA