The Performance of MicrokernelBased Systems

1
The Performance of Microkernel-Based Systems

• L4Linux

2
What is a microkernel

• kernels that only provide address spaces,
  threads, and IPC
• kernel does not handle e.g. the file system or
  interrupts

3
Mircokernel abstraction level

• some researchers feel abstraction level is too
  high
• kernel should map more directly to hardware
• some researchers feel abstraction is too low
• focus on extensibility (Mach)

4
L4

• so called 2nd generation microkernel
• built from scratch as opposed to a developed from
  earlier monolithic kernel approaches (e.g. Mach)

5
L4 essentials

• threads, address spaces and cross-address-space
  communication (IPC)
• other kernel operations e.g. RPC and
  address-space crossing thread migration are built
  up from IPC primitives

6
Address spaces

• recursive construction
• granting, mapping, unmapping
• a page owner can grant or map any of its pages to
  another address space with the receiver
  permission. That page is then accessible to both
  address spaces

7
Address spaces (cont)

• only the grant, map, and unmap are implemented in
  the kernel
• user-level pagers handle page faults

8
Interrupts, exceptions, and traps

• all handled at user level
• interrupts are transformed, by kernel, into IPC
  messages and sent to appropriate user level
  thread
• exceptions and traps are synchronous to
  associated thread and kernel mirrors them to that
  thread

9
Implementation of L4Linux

• develop L4Linux, a linux personality on top of
  the L4 microkernel
• due to time restrictions the linux kernel was not
  fine tuned in L4Linux, so results are only an
  upper bound on the performance penalty

10
L4Linux

• linux 2.0.21 on top of L4
• linux kernel is a user level server
• 100 binary compatible
• modified versions of shared C library libc.so and
  libc.a
• user level trampoline exception
• 14 engineer months and 6500 rewritten lines out
  of a total of 340,000

11
Trampoline

• 100 binary compatible means that a program
  statically linked against the native linux
  library must run, unmodified, on L4Linux
• the trampoline bounces the system-call trap
  that on native linux went into the kernel back
  into the modified shared library on L4Linux
• Microkernel upcalls into user level handler,
  handler than makes an RPC (read, invokes kernel
  again) to OS personality to invoke system call

12
L4Linux (cont)

• L4 maps the entire initial address space to
  kernel server
• single thread in L4, acts as a single virtual
  processor to the linux server
• Linux server occupies a small memory region,
  which utilizes Pentiums segment feature to
  protect its TLB entries, so the TLB always has
  the linux servers translations
  (small-address-space optimization)

13
L4Linux (cont)

• L4 allows user level processes to disable
  interrupts, so uniprocessor version of linux did
  not need modification of critical sections

14
L4Linux (cont)

• interrupt threads have a priority above the
  server itself, so they dont execute concurrently
• signals are forwarded to a co-located signal
  handler inside each user process, since only a
  thread in the same address space can manipulate
  another threads state

15
L4Linux (cont)

• scheduling is mostly done by L4 scheduler
• Four priority levels top half interrupts, bottom
  half interrupts, the linux server, user
  processes. No priority decay.
• so L4 interrupts the linux server in the same way
  the hardware would interrupt a native linux kernel

16
L4Linux (cont)

• user level schedulers can dynamically change
  priority and time slice of any thread?

17
Experiments

• micro- and macro- benchmarks used to compare
  native linux and MkLinux (Mach derived variant)
  to L4Linux
• linux vs L4Linux demonstrates performance penalty
  for using microkernel
• L4Linux vs MkLinux demonstrates influence of
  mircokernel on overall system including the
  influence of colocation
• extensibility experiments
• functionality specialized for L4Linux

18
PerformanceL4Linux, MkLinux and Linux

• microbenchmarks
• getpid L4Linux 2.4 or 3.4 times slower than
  linux MkLinux 3.9 or 28 times slower than
  L4Linux
• lmbench and hbench L4Linux 1 to 3 times slower
  than linux MkLinux 1 to 32 times slower than
  L4Linux

19
Performance (cont)L4Linux, MkLinux and Linux

• macrobenchmarks
• recompiling linux server L4Linux 6-7 slower
  than linux MkLinux 10-20 slower than L4Linux
• AIM multiuser benchmark suite job throughput in
  L4Linux is 7-8 lower than linux MkLinux is
  30-52 lower than L4Linux

20
Conclusions

• At application level there is a 5-10
  performance penalty for using L4Linux vs bare
  linux
• The particular microkernel used matters
• Colocation it secondary to microkernel
  implementation

21
Extensibility

• Can we add services outside L4Linux to improve
  performance by specializing Unix functionality
• Can we improve certain applications by using
  native microkernel mechanisms in addition to the
  classical API
• Can we achieve high performance for
  non-classical, Unix-incompatible systems
  coexisting with L4Linux

22
Pipes and RPC
23
Virtual Memory

• measure user level page fault that maps a page
  from one address space to another (not available
  on Unix)
• measured traps and two different trap, protect,
  unprotect patterns which performed on average 4
  times faster than native linux

24
Cache Partitioning

• User level main-memory manager can coordinate
  with L4 to allocate specific L2 cache pages to
  certain processor
• matrix multiplication example with a four times
  speed-up of worst case performance

25
Possible Alternatives

• Protected Control Transfers
• Grafting