1 / 30

Filesystem Wrapup & Misc

Filesystem Wrapup & Misc. Vivek Pai Princeton University. Who Are These People?. Moral: Andy Warhol’s usually right. Some Auspicious Signs. On Saturday Night Live, a parody of The Rev. Jesse Jackson quotes Hegel’s “thesis, antithesis, synthesis” Not really funny, but hey, it’s Hegel

eman
Télécharger la présentation

Filesystem Wrapup & Misc

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Filesystem Wrapup & Misc Vivek Pai Princeton University

  2. Who Are These People? Moral: Andy Warhol’s usually right

  3. Some Auspicious Signs • On Saturday Night Live, a parody of The Rev. Jesse Jackson quotes Hegel’s “thesis, antithesis, synthesis” • Not really funny, but hey, it’s Hegel • On Sunday, I was flipping channels and saw Kundun • Philip Glass did a pretty good job • See also Thin Blue Line

  4. Thesis – Load Balance Is Good • Imagine using a cluster of machines as a web server instead of a big parallel box Server Incoming requests Server Server

  5. Antithesis – Load Balance Is Bad • Popular files get duplicated in all caches • Effective cache size doesn’t scale with number of nodes • Alternative: locality-based distribution • Problem: Conflict - locality-based determines where request goes, so does load balance

  6. What Happens in Locality-Based? • What if files A, B, C popular? • What if popularity changes? Server A, B, C Incoming requests Server D, E, F Server G, H, I

  7. Previously New Approach Synthesis – Combine Load Balancing with Locality-Based Decisions

  8. What I Did This Past Week • Thursday Night – ended at Tower • Took visitor around Princeton • Friday – Campus • Watched EE keg race, took no sides • Saturday – No partying, just reading • Yes, I know, Winter Formals • Sunday – Fly to DC (bad idea) • Monday – Meeting all day

  9. Observations • I’ve got well-placed students • My 318’ers sing Bon Jovi • Several of us sing Tiffany • Manish manned the mike fairly well • Swope can actually sing • Swope can’t throw

  10. UFS and FFS • The Unix File System is called UFS • It gets named that in retrospect • The 4.2 BSD filesystem was often called FFS – Fast File System • Nowadays, it’s standard, so it’s UFS • But why name it? • Answer: file systems decoupled from OS

  11. VFS – Gaining Flexibility • VFS = Virtual File System • Function pointers to initialize, mount, unmount, map to/from vnode, etc • Vnode = virtual (i)node • Ref counts on disk, in memory • List of clean, dirty blocks in memory • Vnode ops • Open/close, read/write, create/delete (files/dirs) , logical/phys map, etc., etc.

  12. Filesystems Available On My Box • ufs – pre-4.2 Unix • ffs – Fast File System • mfs – Memory-based (temporary) • hpfs – OS/2? Not sure • msdosfs – use DOS disks on Unix • isofs – the format for CD-ROMs • nwfs – Netware? • ntfs – Windows NT • Misc – lots of pseudo-filesystems

  13. Mounting – Mix and Match • Question: what happens if you have multiple disks? • Problem: only one ‘/’ directory • Answer: “mount” disks anywhere you have a directory • Benefit: single directory structure

  14. What Does Mounting Look Like / home var usr bwk vivek willy New disk

  15. Network File Services • Idea: use network to access disk • Benefit: many machine can share disk • Approach: have standard protocol to communicate between client and server client client server client client

  16. Pros & Cons of NFS • Benefits: • Fewer disks to backup • Fewer points of configuration • Allows users to move to any machine • Drawbacks: • DMA fast, network slow • Lots of data traversing network • Possible solution: client caching

  17. Distributed Consistency • Locally cache pages from server • Problem: what happens when one process writes to a shared page? client client server client client

  18. Fun Quotes • “My butt hurts from being thrown on the table” • “I just want everyone to know I’m not drunk. Double E is the best major in the world!” • “I’m a spanker. I spank everyone. They all know that.” • “That outfit doesn’t match. At all.” • “Hey, wait, none of this will affect my grade, will it?”

  19. How To Be A Millionaire • A quick look at OS ideas that have made people rich

  20. Sun – Eliminating Vendor Lockin • Scenario: lots of minicomputer vendors • Many competing operating systems • Unix getting popular, fragmented • Answer: built reasonably-priced workstations using “commodity” components, license OS from AT&T

  21. Auspex – Resources for Speed • Scenario: NFS getting popular, but machines run slowly • Observation: network processing eats CPU time, but it can be parallelized • Approach: build Sun-compatible parallel hardware, license OS from Sun, make modifications to speed NFS

  22. Presto – NVRAM for Disks • Scenario: some workloads slow due to synchronous disk operations • Approach: add hardware with nonvolatile RAM (NVRAM), use that as a stable, writable cache • Also: need to modify OS to understand the existence of this cache

  23. Tandem – High Reliability • Scenario: machine reliability is key, various kinds of failures (even Byzantine) • Approach: use triple-redundancy & voting – easily handles any single failure • Customers – stock markets, particularly

  24. NetApp – A Better Auspex • Scenario: Auspex doing well, but disk bottlenecks become dominant issue • Approach: build appliances (like Auspex) but use a custom OS and custom filesystem • Details – sort of like LFS, but doesn’t compress segments, just fills in holes

  25. Netscape – Friendliness • Scenario: some college kids have made the web big, business don’t like to deal with college kids • Approach: recreate the same things in a corporate environment, give away browser, charge for server • Work: getting decent performance across broad range of OSs

  26. Middleware – Filling in Niches • Scenario: bandwidth is expensive, and many people from a site access the same pages • Background: CERN releases a proxy cache, which can serve “hot” requests locally rather than getting remotely • Opportunity: CERN proxy not being pushed hard, no performance emphasis • Buyer: Network Appliances

  27. Falcon – Caring About Speed • Scenario: proxy servers exist in many places, academics doing a poor job of studying them • Approach: optimize performance on regular OSs, find a buyer with deep pockets who wants to enter market

  28. VMWare – Reinventing VMs • Scenario: machines take space and resources, but we rarely push them • Approach: build a hypervisor to virtualize the x86 architecture • Ugliness: the 386+ was never meant to be virtualized

  29. CacheFlow – A Better NetApp • Scenario: NetApp busy selling fileservers, sales force doesn’t focus on proxy servers • Approach: take similar approach, build proxy servers only

  30. Lessons From All Of This • In retrospect, it always seems easy • Plenty of good ideas available • Sometimes, repetition doesn’t hurt • The window of opportunity still doesn’t seem to have closed • Education pays off • OS has been a “hot” field for a while

More Related