Technology and Operating Systems

Lecture Notes for CS 140
Spring 2019
John Ousterhout

• Many of the basic ideas in operating systems were developed 30-50 years ago, when technology was very different. Are these ideas still relevant today and in the future?
• Technology changes over the last 30 years:
  • CPU speeds: 15 MHz -> 2.5 GHz (167x)
  • Memory size: 8 MBytes -> 16 GBytes (2000x)
  • Disk capacity: 30 MBytes -> 2 TBytes (66667x)
  • Disk transfer rate: 2 MBytes/sec -> 200 MBytes/sec (100x)
  • Network speed: 10 Mbits/sec -> 10 Gbit/sec (1000x)
• The role of paging:
  • When originally proposed (1960's):
    • Disk speed: 80ms latency, 250 KBytes/sec transfer
    • Memory size: 256 KBytes (64 pages)
    • Time to replace all of memory:
      • 6.4 sec (individual page faults)
      • 1 sec (sequential)
  • Today:
    • Disk speed: 10ms latency, 150 MBytes/sec transfer
    • Memory size: 64 GB (16,000,000 pages)
    • Time to replace all of memory:
      • 44 hours (individual page faults)
      • 320 sec (sequential)
  • Can't afford to page something out unless it's going to be idle for a long time.
  • Does paging make sense anymore?
    • Mechanism for incremental loading of processes?
      • Why not just read the entire binary at once?
      • 15 MB of binary takes .1 sec.
    • Safety valve for temporary emergencies?
      • Perhaps, but not much space between "system not paging at all" and "system totally unusable".
    • Maybe NVMs will make it practical again?
  • Virtual memory still be useful:
    • Simplifies physical memory management
    • Allows controlled sharing
    • Memory-mapped files
    • Virtual machines
  • Page size is way too small:
    • Random accesses for replacement too expensive.
    • Not enough TLB coverage.
• TLBs:
  • Haven't kept up with memory sizes
  • 64 entries -> 256 KBytes coverage
  • In mid-80's this was a substantial fraction of memory (8 Mbytes).
  • Today TLBs can only cover a tiny fraction of memory
  • Some TLBs support larger page sizes:
    • 1 MByte
    • 1 GByte
    • But, this complicates kernel memory management.
• Disks:
  • Capacity has increased faster than access time.
  • Can't actually access all the information you can store on disk!
  • Frequently accessed information must move elsewhere
• Storage latency:
  • Disks: 10 ms
  • Flash: 100 microseconds
  • New nonvolatile memories (Intel 3D XPoint): 100-300ns
  • Current software stacks totally unsuitable for nonvolatile memories
  • Will the file abstraction make sense in the future?
• Multi-cores
  • For many years, chip technology improvements allowed processor clock rates to improve rapidly.
  • Unfortunately, faster clock rates mean more power dissipation; power limitations now limit improvements in clock rate.
  • Chip designers are now using technology to put more processors (cores) on a chip.
  • Consequences:
    • All OSes must now be multiprocessor OSes
    • Not clear how to utilize all of these cores: app developers must now write parallel programs?
    • Writing parallel programs is very hard
• Current hot area for OS development: the datacenter
  • Coordinating thousands of machines working together
  • Potential for very low-latency communication.