Measured Performance of Commodity Operating Systems

1
Measured Performance of Commodity Operating
Systems

• Work done at Harvard University
• CSE-597A Presentation
• V.N.Murali

2
Initial Remarks

• Any ideas?
• Windows for WorkgroupsMost expensive due to
  frequent changes in machine mode and system call
  hooks.
• NTmore efficient than above,overhead due to
  microkernel-like design(Win-32 API is a
  user-level server)
• NetBSDMost efficient

3
Components
WINDOWS Segmented memory,Virtual Machine Monitor,Windowing System,Graphics engine WINDOWS NT Executive(IPC,VM,IPC,File system,N/W) Win32 Subsystem(Windowing System) NET BSD Monolithic kernel,X11 Windowing System
4
Device Drivers

• WindowsVxD(virtual Device Drivers) which are
  activated by hardware interrupts,system calls
  and/or entry points.
• NetBSD/dev file is installed.Load a driver and
  access the counter using the device
  file.Configured using ioctl()
• NTDDk is the API for driver development.Loaded
  into the NT executive

5
Microbenchmarks

• Null counter access latency.
• Syscall minimum system call latency
• Exec latency to load and run a program
• Memory access time access times for arrays
  (various sizes)
• File System
• BitBlt Graphics subsystems
• Network

6
Application workloads

• Wish command interpreter for the Tcl language
  (test for GUI subsystem,CPU intensive,very little
  Disk activity)
• Ghostscript
• WWW server
• (ASSUMPTIONSSingle user mode,little or no
  background activity)

7
Metrics

• Cycle counts.
• Instruction counts ? ( Would this be a correct
  metric ?)
• Data read/write references ??(Similar arguments)
• Cache and TLB misses (Icache,Dcache,ITLB,DTLB)

8
RESULTSNull micro

• Least instruction count Windows. Why?
• Maximum instruction count Windows NT.Why?
• Higher CPI Windows.Why?
• Highest cycle counts NT.Why?
• Highest instruction cache miss NT.Why?

9
RESULTSSyscall micro

• Used a simple system call(dup) on UNIX and
  NT.Get_extended_error_info (16bitWin,int21),get_in
  terrupt_vector (32bit,int 21)..
• Overhead for dup in NT and BSD is same as Null
  benchmark.Similar behavior in I-cache misses for
  NT.
• Most efficient in terms of cycle counts? Windows
  32 bit system call?

10
Contd.

• Windows 16 bit is very expensive?Why?
• Hooks?System calls can be intercepted in DOS.Used
  for CDROM drivers,caching etc
• Most EfficientNetBSD and Windows 32 bit.Both NT
  and Windows 16 are expensive.

11
RESULTSExec micro

• NetBSD uses a vfork() and exec() combo while NT
  uses CreateProcess().Windows has a shared address
  space,so nothing is created.
• Most efficient NetBSD static.Why?
• Worst NT dynamic.Why?

12
Contd..

• Windows overhead is higher than NetBSD static and
  lower than either OSs dynamic linkage
• Maximum data reads and writes NT.Why?
• Dynamic linkage incurs a lot of overhead.

13
RESULTSMemory Access time

• Repeated references to arrays of various sizes
  using stride128bytes.
• 8k on chip cache and 256K onboard cache.
• NT uses a deterministic page mapping policy gt
  similar performance upto 256K and smooth
  degradation afterwards
• NetBSD uses a non-determinstic page mapping
  policy.Hence poor performance for gt8k size

14
Contd.

• NTs file system buffer cache is integrated to
  the VM system.This rings a bell doesnt it??
• 64k segment size in Windows limits performance

15
RESULTSFile System Micro

• 3 activities,a)Hit in disk cache b)Access to
  small files c)Creation
• Tested NTFS and FAT32 for Windows NT,FAT 16,32
  for Windows,FFS for NetBSD.
• Least overhead NetBSD
• Most overhead NT/NTFS,FAT 16.

16
Contd..

• Meta data updates NTFS performs the best because
  it logs them,while FFS performs poorly because of
  synchronous writes.

17
RESULTSGraphics micro

• Display an array of pixels repeatedly
• Worst performance . NetBSD.Why?
• NETWORK THROUGHPUT
• Worse code locality for NT gt instruction cache
  and TLB misses.
• However throughput is comparable to
  BSD.Limitation is due to the ethernet.

18
Summary

• Frequent CPU mode changes in Windows expensive
• 64K segment size is a limitation
• Higher instruction counts and cache misses for NT
  ?Why?
• Efficient graphics and relaxed file system
  semantics in NT

19
Application workloads

• WishOverhead includes context switches and
  IPC.Highest overhead in Windows.Why?
• NT is worse than BSD.Why?
• Ghostscript Best performance by NT and Windows
• Web ServerBest results for NetBSD and
  Windows,Worst results for NT.NT is believed to
  have a very poor network implementation.