1. URPC for Shared Memory Multiprocessors Brian Bershad, Thomas E. Anderson, Edward D. Lazowska, and Henry M. Levy ACM TOCS 9 (2), May 1991

2. IPC Performance Efficient IPC - central to OS design: encourages systems decomposition across AS –Failure isolation –Extensibility –Modularity But, performance determines its usability

3. Kernel-based IPC - problems Architectural performance barriers – costs of invoking kernels and processor reallocation (  70% overhead in LRPC) Interaction bet/ kernel-based comm. & high- performance user-level thread mgnt.

4. URPC for SM multiprocessors Solution – eliminate kernel from path –Use SM for data transfer –Take advantage of P already in AS Advantages –Msgs. Sent bet/ AS w/o invoking kernel –Avoid unnecessary P reallocation –When necessary, cost amortized Only P reallocation requires kernel invocation - contrast w/ microkernels!

5. RPC idea and definition Apps/OS service comm. through messages vs. procedure calls  RPC RPC – synchronous lang-level control transfer bet/ programs in disjoints AS whose primary comm mech is a narrow channel Nothing of –narrow channel operations –Processor scheduling mech. interaction w/ data transfer

6. URPC Msgs exchanges bet/ AS using SM User-level thread mgnt integrated w/ user- level msg channel mgnt When a T in a client invokes a procedure in a server T blocks P serves another T in same AS … same on server side User’s view is unchanged

7. Processor reallocation & context switching Context switching – switching P bet/ Ts in same AS (15  sec) Processor reallocation – allocating P to T in another AS (55  sec w/o long-term costs) Costs Changing mapping registers defining virtual AS (immediate) Decide the AS Diminishing benefits from cache and TLB (long-term)

8. Processor reallocation Sometimes necessary Underpowered AS – an AS w/ pending incoming msgs P balances load by reallocating itself Detecting incoming msgs and scheduling T done by low-level T in URPC / P scan for incoming msgs only when idle

9. Example EditorWinMgrFCMgr T1 Call (send/recv WinMgr) T1 Call (send/recv FCMgr) T2 Call (send/recv FCMgr) Recv & process reply T1 Recv & process reply T2 Context switch – terminate T2 Context switch Recv & process reply T1 Context switch – terminate T1 Processor realloc Time

10. Design rationale Data transfer –SM msg channels give same safety guarantees Processor reallocation Optimistic Client has other work to do (thread or incoming msgs) Server has or soon will have a P to use When wrong: P reallocation

11. Design rationale Threads High-performance T mgnt necessary for fine- grained parallel programs This can only be done at user-level Due to close interaction – comm has to be done at user level Heavyweight / Middleweight / Lightweight threads Lightweight threads  user-level comm.

12. Performance evaluation TestURPCTaos Procedure call77 Fork431192 Fork;Join1021574 Yield3757

13. Performance evaluation ComponentClientServer Send1813 Poll66 Receive109 Dispatch2025 Total5453