1. Improving IPC by Kernel Design Jochen Liedtke Proceeding of the 14 th ACM Symposium on Operating Systems Principles Asheville, North Carolina 1993

2. The Performance of u-Kernel-Based Systems H. Haertig, M. Hohmuth, J. Liedtke, S. Schoenberg, J. Wolter Proceedings of the 16 th Symposium on Operating Systems Principles October 1997, pp. 66-77

3. Jochen Liedtke (1953 – 2001) 1977 – Diploma in Mathematics from University of Beilefeld. 1984 – Moved to GMD (German National Research Center). Build L3. Known for overcoming ipc performance hurdles. 1996 – IBM T.J Watson Research Center. Developed L4, a 12kb second generation microkernel.

4. The IPC Dilemma IPC is a core paradigm of u-kernel architectures Most IPC implementations perform poorly Really fast message passing systems are needed to run device drivers and other performance critical components at the user-level. Result: programmers circumvent IPC, co-locating device drivers in the kernel and defeating the main purpose of the microkernel architecture

5. What to Do? Optimize IPC performance above all else! Results: L3 and L4: second-generation micro- kernel based operating systems Many clever optimizations, but no single “silver bullet”

6. Summary of Techniques Seventeen Total

7. Standard System Calls (Send/Recv) send ( ); System call, Enter kernel Exit kernel Client (Sender)Server (Receiver) receive ( );System call, Enter kernel Exit kernel send ( ); System call, Enter kernel Exit kernel receive ( ); System call, Enter kernel Exit kernel Client is not Blocked Kernel entered/exited four times per call!

8. New Call/Response-based System Calls call ( ); System call, Enter kernel Allocate CPU to Server Suspend Re allocate CPU to Client Exit kernel Client (Sender)Server (Receiver) Resume from being suspended Exit kernel reply_and_recv_next ( ); Enter kernel Send Reply Wait for next message handle message Special system calls for RPC-style interaction Kernel entered and exited only twice per call! reply_and_recv_next ( );

9. Complex Message Structure Batching IPC Combine a sequence of send operations into a single operation by supporting complex messages Benefit: reduces number of sends.

10. Direct Transfer by Temporary Mapping Naïve message transfer: copy from sender to kernel then from kernel to receiver Optimizing transfer by sharing memory between sender and receiver is not secure L3 supports single-copy transfers by temporarily mapping a communication window into the sender.

11. Scheduling Conventionally, ipc operations call or reply & receive require scheduling actions: –Delete sending thread from the ready queue. –Insert sending thread into the waiting queue –Delete the receiving thread from the waiting queue. –Insert receiving thread into the ready queue. These operations, together with 4 expected TLB misses will take at least 1.2 us (23%T).

12. Solution, Lazy Scheduling Don’t bother updating the scheduler queues! Instead, delay the movement of threads among queues until the queues are queried. Why? –A sending thread that blocks will soon unblock again, and maybe nobody will ever notice that it blocked Lazy scheduling is achieved by setting state flags (ready / waiting) in the Thread Control Blocks

13. Pass Short Messages in Registers Most messages are very short, 8 bytes (plus 8 bytes of sender id) –Eg. ack/error replies from device drivers or hardware initiated interrupt messages. Transfer short messages via cpu registers. Performance gain of 2.4 us or 48%T.

14. Impact on IPC Performance For an eight byte message, ipc time for L3 is 5.2 us compared to 115 us for Mach, a 22 fold improvement. For large message (4K) a 3 fold improvement is seen.

15. Relative Importance of Techniques Quantifiable impact of techniques –49% means that that removing that item would increase ipc time by 49%.

16. OS and Application-Level Performance

17. OS-Level Performance

18. Application-Level Performance

19. Conclusion Use a synergistic approach to improve IPC performance –A thorough understanding of hardware/software interaction is required –no “silver bullet” IPC performance can be improved by a factor of 10 … but even so, a micro-kernel-based OS will not be as fast as an equivalent monolithic OS –L4-based Linux outperforms Mach-based Linux, but not monolithic Linux