Transcript ppt - Computer Science

Exokernel: An Operating System
Architecture for Application-Level
Resource Management
Dawson R. Engler, M. Frans Kaashoek, and James O'Tool Jr.
M.I.T Laboratory for Computer Science
Cambridge, MA 02139, U.S.A
Presented by Jennifer Minor
What is a Kernel?
Definition from wiktionary.org:


The core, center, or essence of an
object or system.
(computing) The central part of many
computer operating systems which
manages the system's resources and
the communication between
hardware and software components.
So what is an Exokernel?
A Monolithic Kernel is...




All operating system services run in
kernel mode.
Single address space.
High level abstractions given to
application.
Must support a wide range of
applications.

Slow to change.

System Calls are expensive.
A Microkernel is...





Separate mechanism from policy.
Only lower level mechanisms are
supported in kernel mode. (Address
space management, scheduling and
basic IPC)
Policies are implemented in user
level which are easier to change.
Kernel must protect servers from
each other.
Good protection but has to use IPC to
communicate.
So an Exokernel is...




Similar to microkernel in that only
minimum functionality is in the kernel.
Unlike the microkernel it exports
hardware resources rather than
emulating them.
Physical resources are safely allocated to
the application, where it can be
managed.
All abstractions are implemented in
application-level or as part of a library
OS that is part of the application address
space.
Exokernel Architecture
Goal: Separate protection from management.
1. Low-level interface:
Provide simple and efficient primitives.
2. Multiplex resources:
Securely and fine-grained.
3. Limit management to protection:
Protect without specific usage knowledge of resource.
4. Export hardware resources:
Expose hardware and kernel data structures.
5. Notify Application:
Event notifications and visible resource revocation.
Exporting Resources Securely
1. Secure Bindings

Hardware mechanisms

Software caching

Downloading application code
2. Visible Resource Revocation

Application level guided deallocation

Application specific knowledge of state needed to be saved

Application notification that resources are scarce
3. Abort Protocol

Mechanism for kernel to force-ably take back resources.

Still notifies application after the fact.
Aegis: an Exokernel

Processor Time Slicing

Represents CPU as a linear vector partitioned time
slices that can be allocated by the application.



STLB
A large software TLB is over the hardware TLB
and can be used on a cache miss to map address.
Dynamic Code Generation
Creation of executable code at runtime. Used
by the network subsystem to download filters
for demultiplexing messages.
Processor Environments
Structures that store information needed to deliver
events to applications. (Upcalls)

Holds application data and code in memory.
Also allows each application a small number of
pinned virtual addresses.
Timer Interrupts
Denote the beginning and end of a time slice to the
user-level code where scheduler activations can be
implemented.
Guaranteed Mappings

Protected Control Transfers
Changes the program counter to callee, donates
current time slice to callee's processor
environment and switches to the callee's
context.
User level efficient IPC abstraction can easily be
built on top of PCT's.
Aegis: Events
Four Types: Exceptions, Interrupts, Protect Control Transfers and Address Translations
Event Handler Contexts Include:



Program counter to jump to on event.
Memory location to save registers.
Additional status registers are needed for timer interrupts and tlb misses.
What happens on a hardware exception?



Aegis saves three scratch registers into the “save-area”.
Loads the exception program counter, the last virtual address translation and cause.
Performs a indirect jump into an applications-specified program counter.
Note: After handling the exception the application can resume execution without going back to the kernel.
Special event handlers have to be defined for start-time-slice, end-time-slice,
asynchronous control transfers, and synchronous control transfers.
Aegis: Performance
Machine
OS
DEC2100
Ultrix
DEC2100
DEC3100
DEC3100
Aegis
Ultrix
Aegis
DEC5000
DEC5000
Ultrix
Aegis
Procedure call Syscall (getpid)
0.57
0.56
0.42
32.2
3.2 / 4.7
33.7
0.42
0.28
0.28
2.9 / 3.5
21.3
1.6 / 2.3
Why is performance so much better on Aegis?


Kernel data structures are not mapped. No need to worry
about a interrupted TLB miss.
Two paths for system calls, one for calls that require a
stack and a second for ones that do.
ExOS: a Library Operating System
Implements traditional operating system abstractions at the
application level, since it runs in the applications address space.

Fault Isolation



Each application runs in it's own address space.
Efficient


No protection domain crossing to manage
resources after they have been allocated.

System calls are near procedure call speed.
Extensible

Policies can be altered at application level.
Built on top of protected control transfers.
Virtual Memory
Using low-level hardware abstractions
ExOS provides a rudimentary VM system.


IPC abstraction


Remote Communications
Downloading code into the kernel allows
the demultiplexing of the messages without
a context switch.

ExOS: IPC Performance


Machine
OS
pipe
pipe'
shm
lrpc
DEC2100
Ultrix
326.0
n/a
187.0
n/a
DEC2100
Aegis
30.9
24.8
12.4
13.9
DEC3100
Ultrix
243.0
n/a
139.0
n/a
DEC3100
Aegis
22.6
18.6
9.3
10.4
DEC5000
Ultrix
199.0
n/a
118.0
n/a
DEC5000
Aegis
14.2
10.7
5.7
6.3
ExOS built a lrpc abstraction on top of the low-level protected
procedure call interface given by Aegis.
Ultrix does not currently have a lrpc implementation to add new
functionality it would need to build on top of one of the existing
high-level abstractions such pipes.
ExOS: VM Performance
Machine
OS
dirty
prot1
prot100
unprot100
trap
appel1
appel2
DEC2100
Ultrix
n/a
51.6
175.0
175.0
240.0
383.0
335.0
DEC2100
Aegis
17.5
32.5
213.0
275.0
13.9
74.4
45.9
DEC3100
Ultrix
n/a
39.0
133.0
133.0
185.0
302.0
267.0
DEC3100
Aegis
13.1
24.4
156.0
206.0
10.1
55.0
34.0
DEC5000
Ultrix
n/a
32.0
102.0
102.0
161.0
262.0
232.0
DEC5000
Aegis
9.8
16.9
109.0
143.0
4.8
34.0
22.0


Kernel transitions can be eliminated by implementing abstractions at
application level.
Application-level software can implement functionality that is frequently not
provided by traditional operating system.
ExOS: Application-Specific Safe Handlers
ASH: Untrusted applicationlevel message-handlers that
are downloaded into the
kernel, made safe with code
inspection and sand boxing.
Roundtrip Latency (microseconds)
3500
3250
ExOS without ASH
3000
ExOS with ASH
2750
2500
2250
• Reduces intermediate
copies of message.
2000
1750
1500
• Can integrate check
summing in transfer
mechanism.
1250
1000
750
500
250
0
1
2
3
4
5
6
7
Number of Processes
8
9
10
• Low-latency message
replies
• Control initiation
Why are Exokernels important?
Fixed high level abstractions hurt application performance
 Fixed high level abstractions hide information
 Fixed high level abstractions limit the functionality

"Because all applications must share the core abstractions, changes to core abstractions
occur rarely, if ever. This is perhaps why few good ideas from the last decade of operating
systems research have been adopted into widespread use. What operating systems support
scheduler activations [3], multiple protection domains within a single address space [10],
efficient IPC [29], or efficient and flexible virtual memory
primitives [4, 21, 25]?”
Exokernel Design Proves:
Resources can be securely partitioned with low overhead
 Low-level interfaces and exposed kernel data structure
can produce efficient implementation due to simplicity
 Downloadable application code into the kernel increase
performance and responsiveness
 Library Operating Systems provide extensible and
customizable services at application level.

References
MIT Exokernel Operating System http://pdos.csail.mit.edu/exo.html
 Wikipedia: Exokernel http://en.wikipedia.org/wiki/Exokernel
 Wikipedia: Kernel (computer science)
http://en.wikipedia.org/wiki/Kernel_%28computer_science%29
 Wikipedia: MicroKernel http://en.wikipedia.org/wiki/Microkernel
 Wikipedia: Monolithic Kernel http://en.wikipedia.org/wiki/Monolithic_kernel
Wiktionary: kernel http://en.wiktionary.org/wiki/kernel
