CSE221 - lec07: Distributed OSes: Distributed V & Sprite

07 - Distributed OSes: Distributed V & Sprite

The Distributed V Kernel and its Performance for Diskless Workstations

Basic model for diskless workstations

• Diskless workstations, small local disk

• Access files stored in file servers via local network

Claims

• Diskless workstations is comparable to those use local disks in terms of performance

• General purpose IPC with messages is sufficient for communication. (No specialized protocol is needed)

V Inter-Process Communication

• Synchronous request / response messages
• slow & inefficient (-)
• simpler (+)
• easier buffer management (+)
• familiar programming model (+)

• Small fixed-size messages
• used to pass control messages (32 bytes)
• via Send / Reply
• easier buffer management (+)
• sufficient for lots of messages (+)
• inefficient network usage (-)

• Separate data transfer mechanism
• used to pass larger pieces of data
• via MoveTo / MoveFrom for large-size data
• via ReceiveWithSeg / ReplyWithSeg for medium-size data

Performance Optimization

• use raw ethernet

Network Penalty

• minimum time to transfer certain amounts of data via network

• a lower bound of cost to go distributed.

Experiment 1: Kernel Performance Evaluation

• goal: measure the overhead of IPC primitive, comparing the local v.s. remote

• remote add few ms of overhead

• disk access latency is about 20 ms, much larger than the overhead

Experiment 2: File Access

• three forms of file acess: random file page access, sequential file reading, and program loading.

• Streaming:
• form 1: streaming over network (not used in V)
• form 2: prefetching and batch write (do apply this in V)
• arguments: don’t need streaming protocol over network
• I/O bound applications: disk is the bottleneck
• not I/O bound: improvement is small, not worth optimizing

Experiment Summary

• contains both micro + macro benchmark, measuring both primitive operations and end-to-end system performance.

• use experiment results to support the advocate for diskless workstations: gain additional benefits without losing many performance

Summary

• The notion of diskless workstations

• arguments based on performance

The Sprite Network Operating System

Tech Trends

• large memory → more caching

• higher performance network → workstations connected via network
• unreliablibity of underlying network
• consistency
• slower
• manage / administering is hard
• information sharing → addressed by single global namespace in this paper
• idle machine → addressed by process migration in this paper

• multi-processors
• more sharing & parallelism: shared_fork()
• OS organization: subsystems + monitors

Caching

• Consistency problem: when the cache contents in different clients differ
• Two types of consistency problems
• sequential write-sharing: one write and close, then another open. → resolve using version numbers
• concurrent write-sharing: multiple clients modify contents concurrently → disable client cache

Physical Memory Allocation

• Virtual Memory and File System has conflict use of physical memory: vm wants more to improve memory performance, fs wants more to improve file access performance.

• can be resolved by vm-fs negotiation: dynamically adjust based on statistics information from vm and fs.

Process Migration

• goal: transparency (the user don’t know whether the process is migrated or running locally)

• three step: freeze - transfer - unfreeze

• handle machine dependent kernel calls:
• forward to home node kernel (where the process is created)

• Today: no process migration, more VM / container migration

Summary

• file caching protocol to resolve consistency problem

• tech trends affect system design and research problems