1. Fail-safe Communication Layer for DisplayWall Yuqun Chen

2. DisplayWall Software Architecture render node render node render node master node Logical Network 1 1 1 1 22 2 33 1 3 3 Command Broadcast Synchronization Data Exchange

3. Motivation Complex communication patterns –programming the DisplayWall is difficult BitBlt operation in Virtual Display Driver Nodes and network links do fail –larger system is more likely to fail –OS may not be stable under high-load –applications have bugs

4. Design Goal [1] Eases writing distributed applications on DisplayWall –supports some form of group concept –multicast or broadcast –no need to manage pair-wise connections

5. Design Goals[2] An API for designing fail-safe DisplayWall apps –tolerate independent failures and recovery of render nodes –failures at the application master nodes are considered catastrophic –certain bugs cause all render nodes to fail may or may not be able to deal with this

6. What’s different with others? API-wise –mostly broadcast –some pair-wise exchange Fault-tolerance wise –real-time characteristic cannot wait for too long –OK to lose certain messages dropping a few frames is OK

7. Some Requirements Simple abstraction –users shouldn’t deal with pair-wise connections Realizable on a variety of platforms –with and without programmable NI Support storage and retrieval of application- dependent states (soft states) Synchronized Clocks and Barriers

8. Communication Patterns Command/Data Delivery –from master to some render nodes –broadcast in nature Data exchange –among render nodes, e.g., bitblt and v-tiling –pair-wise in nature Synchronization: clock and barriers –low-latency

9. Outline Communication patterns Soft States API issues

10. Command Broadcast Used by all applications –VDD, OpenGL, ImageViewer Issues: –efficiency –dynamic membership: live/dead nodes, overlapped windows –delivery semantics best-effort, guaranteed, or sloppy

11. Guaranteed or Best-effort? Guaranteed delivery implies –re-configurable (logical) topology for delivering data –loss-less delivery to all nodes an intermediate node may fail right after delivery –the up-stream has to keep all the data for retransmission Best-effort –as far as current topology allows –or limited flexibility

12. Transactions? Each broadcast is treated as a transaction –keeps the data until the transaction commits Transactions are asynchronous –one doesn’t wait till the previous commits –but, they are applied in order Applied transactions mark the state at each node

13. Fail-safe Broadcast 1234 1234 1234

14. recv(msg) send msg to children c1, c2, …, cn if child c failed then recompute c’s subtree without c send to the new child cc

15. Questions How to detect failure in a timely fashion? –global solution the leaves send the acks to a master the master forces a global reconfiguration after a timeout –a local solution period positive ACK to the parent

16. Data Exchange Pair-wise sends and recvs High-level issues: –avoid the recv to get stuck by a failing sender timeout or period “I am alive” ack Implementation issues: –neighborhood communication? probably sufficient for most apps except for load balancing

17. Synchronization Barrier –a special form of broadcast –mostly used for global frame-buffer swap Clock synchronization –can be used to reduce the frequency of barriers –e.g., MPEG playback according to local clock what if it misses the deadline?

18. Unification Transaction-based messages implement broadcast –implements some form of global ordering Message passing for pair-wise data exchange –key is to detect the failures quickly

19. Example 1: VDD BitBlt BitBlt: calculate data to send calculate data to recv read frame buffer send data to some nodes if failure then take note recv data from some nodes if failure then use default image

20. Failure Recovery What happens when a failed render node comes back up? –put itself into the broadcast tree –re-establish peer-to-peer message connection hopefully all hidden –it has to bring its states up to date highly application dependent

21. Soft States Example: OpenGL Display List –each list consists a series of GL commands must be re-executed to make the list meaningful Textures –may be bound to a texture name

22. Soft States API Tagged Safe Memory –a chunk of memory replicated on all nodes –tagged and ordered by an ID –associate a recovery handle/function to it Operations: create, insert, and delete Upon recovery: –retrieve all “live” chunks and apply the handles in order

23. Example 2: VDD recovery Re-establish connections between nodes Restore States: –cached bitmaps from other running nodes –fonts and brushes from other nodes Or, force a re-draw from the master nodes

24. Messages? Transactions can be implemented on top of messages –add a transaction ID for each message People are familiar with message passing –as opposed to remote memory access you have to manage memory, pretty message What about copy avoidance?

25. Communication API Message Passing Interface –send(node, type,data), recv(node, type, data) –only need to specify a remote node id –connection is hidden from the API –copy can be avoided by returning a buffer pointer instead of filling the user buffer –very close to sockets API but more flexible

26. Copy Avoiding Message Passing Trivial –just return a pointer to the buffer Remote memory semantics not necessary –very rare to update remote data structure –memory copy isn’t that bad ( > 200 MB/sec) What about remote bitblt? –the only missing part is peer-to-peer which we can’t do any way

27. Copy Avoiding Messages CoreLogic global buffer NIC Graphics msg hdr new msg recv(msg)