1. Presented by Neha Agrawal
User-Level Inter process Communication for Shared Memory Multiprocessors
Brian N. Bershad, Thomas E. Anderson, Edward D. Lazowska, and Henry M. Levy
Presented by Neha Agrawal

2. Outline Motivation for RPC Motivation for URPC URPC concept
Components of URPC
Processor Re-allocation
Data Transfer
Thread management
Performance
Conclusion
2019/2/22

3. Motivation for RPC Importance of IPC
Encourages system decomposition across address space boundaries
Failure Isolation
Extensibility
Modularity
Message Passing, a different paradigm
RPC interface above messages
2019/2/22

4. Outline Motivation for RPC Motivation for URPC URPC concept
Components of URPC
Processor Re-allocation
Data Transfer
Thread management
Performance
Conclusion
2019/2/22

5. Motivation for URPC Problems with Kernel based approach
Architectural performance barriers
LRPC: kernel mediates every cross address space call
Kernel based communication hampering high performance user-level threads
Result: Coalesce Weekly related sub-systems
Trading major benefits
2019/2/22

6. Outline Motivation for RPC Motivation for URPC URPC concept
Components of URPC
Processor Re-allocation
Data Transfer
Thread management
Performance
Conclusion
2019/2/22

7. URPC Concept
URPC (User-level Remote Procedure Call) designed for shared memory multiprocessor
User-level management of cross-address space communication
Uses shared memory to send messages directly between address spaces
Avoids unnecessary Processor Reallocation
2019/2/22

8. URPC Concept Division of responsibility Kernel required ONLY
Processor Reallocation
Thread management
Data transfer
Kernel required ONLY
for processor
reallocation.
Exploits RPC definition
Operation of channels of communication
processor reallocation mechanisms
2019/2/22

9. Outline Motivation for RPC Motivation for URPC URPC concept
Components of URPC
Processor Re-allocation
Data Transfer
Thread management
Performance
Conclusion
2019/2/22

10. Processor Reallocation : Why should it be avoided ?
High Immediate costs
Scheduling costs
Cost to update the virtual memory mapping registers
Cost to transfer the processor between address spaces
High Long-term costs
Locality shift of address space
Results in diminished performance of
Cache
TLB
2019/2/22

11. Processor reallocation
Optimistic reallocation policy assumes
The client has more work to do
Inexpensive context switch
Server will have a processor for service
Exploits parallelism
Amortize cost of single processor reallocation
Advantages over
Handoff scheduling
Logically centralized resources
2019/2/22

12. URPC timeline
2019/2/22

13. Why Optimistic approach will not always hold ?
Issues :
Single threaded applications
Real time applications
High latency I/O operations
Priority Invocations
URPC Solution :
Force processor reallocation
2019/2/22

14. The Kernel handles Processor Reallocation
URPC provides Processor.Donate call
When?
Underpowered address space discovered
What parameters ?
Identity of new address space
How ?
Through the kernel to the new address space
Identity of donar made known to the receiver
2019/2/22

15. Voluntary Return of Processors : The Policy
Return the client’s processor when
all outstanding messages from the client processed or
The client has become ‘underpowered’
Policy cannot be ENFORCED
URPC deals with Load balancing only for communicating applications
Preemptive, work conserving policies required to avoid STARVATION
No need for global processor re-allocator
2019/2/22

16. Outline Motivation for RPC Motivation for URPC URPC concept
Components of URPC
Processor Re-allocation
Data Transfer
Thread management
Performance
Conclusion
2019/2/22

17. SAFE Data Transfer using shared memory
Traditional RPC
Clients and servers can overload each other
Need of higher level protocol
Copying Data in and by the Kernel
Unnecessary and insufficient for safety
URPC approach
Use of pair-wise shared memory
Mapped during binding phase
Stubs responsible for safety
2019/2/22

18. Controlling Channel Access
Bidirectional shared memory queue
With a test and set lock on either end
Lock must be NON SPINNING
The protocol
If the lock is free acquire it, else do something else
The Rationale
Sender
Receiver
2019/2/22

19. Outline Motivation for RPC Motivation for URPC URPC concept
Components of URPC
Processor Re-allocation
Data Transfer
Thread management
Performance
Conclusion
2019/2/22

20. Thread management Interaction between : Authors argue :
Send/Receive of MESSAGES &
Start/Stop of Threads
Authors argue :
High performance thread management facilities
Needed for fine-grained parallel programs
Can be provided only at the user level
Have strong interaction with communication
Hence, implement together at user level
2019/2/22

21. Concurrent programming and thread management
Problems with Kernel level threads
Kernel deals with thread management
Expensive kernel traps
General interface for all applications
Avoids customization for applications
Kernel needs to ensure its protection
Adds to the expense of kernel traps
User level threads attacks these costs
Two level scheduling
2019/2/22

22. Threads and communication at two different levels!
Threads at user level and communication at kernel level
Loss of Performance benefits of
Inexpensive scheduling
Context switching
Synchronization would be implemented and executed at two levels
Advantage of both functions implemented at user level
2019/2/22

23. Summarizing the Rationale
Advantages of a user-level approach to communication
Minimal involvement of the kernel
Avoid unnecessarily reallocation of processors between address spaces.
Amortization of the processor reallocation cost.
Integration of user-level communication with user-level thread management facilities for parallelism and performance.
2019/2/22

24. Outline Motivation for RPC Motivation for URPC URPC concept
Components of URPC
Processor Re-allocation
Data Transfer
Thread management
Performance
Conclusion
2019/2/22

25. URPC Performance with no processor reallocation required
Work done during URPC
Send
Poll
Receive
Dispatch
2019/2/22

26. Call latency and Throughput
Both are Load dependant
Both depend on
Number of Client Processors (C)
Number of Server Processors (S)
Number of runnable threads in the client’s Address Space (T)
Call Latency: Time from which a thread calls into the stub until control returns from the stub
2019/2/22

27. Call Latency
Latency is proportional to the number of threads per cpu’s
increases when T > C+S
Pure Latency: 93 sec, S=T=C=1
2019/2/22

28. Throughput Improves until T > C+S Does not doubles Contention
2019/2/22

29. LRPC Vs URPC LRPC uses pair-wise mapped shared memory like UPRC
LRPC uses kernel level threads
hence URPC outperforms LRPC expect for
The case S=0 : LRPC outperforms
Though both LRPC and URPC require 2 processor reallocations and two kernel invocations
Latency
NULL URPC 375 microseconds
NULL LRPC 157 microseconds
2019/2/22

30. Why LRPC wins the S=0 case?
Processor reallocation in URPC based on LRPC
URPC integrated with two level scheduling
Is there an idle processor ? and
is there an underpowered address space to which it can be reallocated ?
2019/2/22

31. Outline Motivation for RPC Motivation for URPC URPC concept
Components of URPC
Processor Re-allocation
Data Transfer
Thread management
Performance
Conclusion
2019/2/22

32. Conclusion
Moving the Communication layer from the kernel to user space can be of advantage
SMMP need specially designed OS’s so that the its processing power can be exploited
URPC is designed for SMMP directly leading to major performance improvements.
2019/2/22