kit – the cooperation of forschungszentrum karlsruhe gmbh and universität karlsruhe (th) system...
TRANSCRIPT
KIT – The cooperation of Forschungszentrum Karlsruhe GmbH and Universität Karlsruhe (TH)
SYSTEM ARCHITECTURE GROUP, KARLSRUHE UNIVERSITY
A Microkernel API for Fine-grained Decomposition
Sebastian Reichelt, Jan Stoess, Frank Bellosa
Component model
Interfaces
Microkernel
OS Decomposition
Sebastian Reichelt, Jan Stoess, Frank BellosaA Microkernel API for Fine-grained Decomposition
2/11
Kernel
Packets
Microkernel
ThreadsAddress spaces
IPC
ThreadsAddress spaces
IPC
TCP/IP stack
Device driver
ThreadsThreads
IPC
Address spaces
Data location?Data transfer?
Split state?System
knowledge?
Ease of development?
Reuse?Granularity?
Microkernel
Our Approach
Sebastian Reichelt, Jan Stoess, Frank BellosaA Microkernel API for Fine-grained Decomposition
3/11
Packets
Microkernel
Component model
TCP/IP stack
Device driver
Interfaces
Function calls
Servers
Local objects
Function calls
Local objects
References
Servers
References
Component model
Data location?Data transfer?
Split state?System
knowledge?
Designing for Decomposition
Sebastian Reichelt, Jan Stoess, Frank BellosaA Microkernel API for Fine-grained Decomposition
4/11
Server
Existing microkernels:Microkernel
ThreadsAddress spaces
IPCCreate thread
Send IPC
Receive IPC
TaskTask
Task Task
Server
Our microkernel: Component model
ServersLocal objects
ReferencesProvided interface
Required interfaces
Threading options
Component Component
ComponentComponent
Reference Management
Sebastian Reichelt, Jan Stoess, Frank BellosaA Microkernel API for Fine-grained Decomposition
5/11
Microkernel
Reference Management
Sebastian Reichelt, Jan Stoess, Frank BellosaA Microkernel API for Fine-grained Decomposition
6/11
openconnection
TCP/IP stack
TCP connection
send packet
Client
Microkernel
Prototype
Multi-server OS (~4 man-months)Basic servers (memory, scheduling, ...)
Drivers (standard PC hardware, PCI, Ethernet)
Decomposed TCP/IP stack
Linux ABI implementation (partial)
Kernel (~2 man-months)IA-32 implementation
Servers in kernel mode
Sebastian Reichelt, Jan Stoess, Frank BellosaA Microkernel API for Fine-grained Decomposition
7/11
Linux
lwIP
Evaluation: granularity, reuse, performance
Granularity
Sebastian Reichelt, Jan Stoess, Frank BellosaA Microkernel API for Fine-grained Decomposition
8/11
Minimum Maximum Average
Lines of code 27 1363 306
Bytes 280 13588 2907
Linux ABI layer
As fine-granular as desired
Size of a server:
Reuse
Sebastian Reichelt, Jan Stoess, Frank BellosaA Microkernel API for Fine-grained Decomposition
9/11
RTL-8139 driver ported from Linux (~1 man-month):
lines of code
e.g. calls via interfacesinstead of directly
queue-basedpacket processing
Reuse possible with little effort
Performance
Sebastian Reichelt, Jan Stoess, Frank BellosaA Microkernel API for Fine-grained Decomposition
10/11
Operation Pentium 4 Core 2
Cross-component call/return (kernel mode)
46 23
Function call/return 11 9
cycles
Operation Pentium 4 Core 2
Cross-component call/return (inter-AS)
1965 771
L4Ka round-trip IPC 2262 1052
Low performance overhead
Conclusion
Microkernel APIs enforce system structureExisting APIs distributed systems
Development, reuse, and granularity problems
Our approach: component model as APIMore general abstractions related to decomposition(servers, interfaces, calls, references, local objects, ...)
Derived microkernel design
Networking-capable prototype OS
Good granularity and reuse, promising performance
Sebastian Reichelt, Jan Stoess, Frank BellosaA Microkernel API for Fine-grained Decomposition
11/11
http://oosys.sourceforge.net/
Demo today