Caching Caching Andrew Security Andrew Scale and Performance

Caching

Caching • Andrew Security • Andrew Scale and Performance • Sprite Performance

Andrew File System

Sprite

Network File System

Andrew File System • AFS, AFS 2, Coda • 1983 to present, Satya its champion • Ideas spread to other systems, NT

Security Terms • Release, Modification, Denial of Service • Mutual suspicion, Modification, Conservation, Confinement, Initialization • Identification, Authentication, Privacy, Nonrepudiation

System Components Authentication Server Vice Virtue Secure Servers Protected Workstations Venus Virtual File System

Andrew Encryption • • DES - Private Keys E[msg, key], D[msg, key] Local copy of secret key Exchange of keys doesn’t scale – Web of trust extends to lots of servers – Pair wise keys unwieldy

Andrew Authentication • Username sent in the clear • Random number exchange – E[X, key] sent to server (Vice) – D[E[X, key] = X – E[X+1, key] to client (Venus) • BIND exchanges session keys

Authentication Tokens • Description of the user • ID, timestamp valid/invalid • Used to coordinate what should be available from Vice (server) to Virtue (client)

Access Control • Hierarchical groups – Project/shared accounts discouraged • • • Positive/Negative Rights U(+) — U(-) VMS linear list & rights IDs Prolog engine in NT Netware has better admin feedback

Resource Usage • Network not an issue – Distributed DOS ‘hard’ • Server High Water Mark – Violations by SU programs tolerated – Daemon processes given ‘stem’ accnt • Workstations not an issue – User files in Vice

Other Security Issues • XOR for session encryption • PC support via special server • Diskless workstations avoided

Enhancements • Cells (NT Domains) • Kerberos • Protection Server for user administration

Sprite Components Client Cache Local Disk Server Cache Server Disk

Sprite Design • Cache in client and server RAM • Kernel file system modification – Affects system/paging and user files • Cache size negotiated with VM • Delayed 30 s write-back – Called ‘laissez-faire’ by Andrew

NFS Comparison • Presumed optimized • RPC access semantics – NFS uses UDP, others TCP • Sprite targeting 100+ nodes • Andrew targeting 5, 000+ nodes

Andrew Scale and Performance • Dedicated server process per client • Directory redirection for content • Whole file copy in cache

Problems already… • Context switching in server • TCP connection overhead – Session done by kernel • Painful to move parts of VFS to other servers – Volume abstraction fixed this later

Cache Management • • Write on close No concurrent write Versioning User level • • Delayed write Cache disabled Versioning Kernel level

Function Distribution • • Test. Auth - validate cache Get. File. Stat - file status Fetch - server to client Store - client to server 61. 7% 26. 8% 4. 0% 2. 1%

Performance Improvements • • Virtue caches directory Local copy assumed correct File id’s, not names, exchanged Lightweight Processes (LWP) – Context data record on server

Andrew Benchmarks

Sprite Throughput

Sprite Benchmarks

Sprite Benchmarks

Cache Impact - Client

Cache Impact - Server

Cache Impact - Net

Comparison

General Considerations • 17 -20% slower than local • Server bottleneck • Scan for files and read almost all local • 6 -8 x faster vs no cache • Server cache extends local cache • Remote paging fast as local disk! • 5 x users/server

Fini