Transcript ppt
ECE 526 – Network Processing Systems Design Software-based Protocol Processing Chapter 7: D. E. Comer Goal • Understand how the network processing protocol stack ─ Implemented ─ Executed ─ Organized in software based network processing systems. Ning Weng ECE 526 2 Outline • Protocol software on conventional processor ─ Possible implementations ─ Processing priority controlling mechanisms • Software interrupts • Software threads ─ Organization of software for layered systems • Summary Ning Weng ECE 526 3 Protocols on Processor • High performance network processing systems ─ Protocols with efficient data structures and algorithms ─ Powerful processor ─ Efficient implementation of protocols on processor • Three possible implementations ─ In an application program • Possible but not practical due lacking of speed ─ In an operating system kernel • • • • Software system managing memory and I/O devices Kernel resident in memory all time when processor on Separate address space, highest privilege Containing device driver and controlling device operations ─ In an embedded system • Programmable hardware device dedicated for special tasks Ning Weng ECE 526 4 Implementation Comparison Easy of Programming High performance Application domain Application Program Easy poor Less performance demanding OS Kernel Most difficult good Higher performance and shared by applications Embedded Possibly difficult good Stand-alone device: bridge Ning Weng ECE 526 5 Data movement within protocols Ning Weng ECE 526 6 Interrupts • What is interrupt? ─ Event signal to OS which results in context switch • Interrupt mechanism ─ ─ ─ ─ Operating asynchronously Saving current processing state Changing processor status Branches to specified locations • Types of interrupt ─ Hardware interrupt: raised by device ─ Software interrupt: raised by executing program • Protocol stack operating as software interrupt ─ When packet arrives, hardware interrupts ─ Device drive raises software interrupt ─ Interrupts cleared and protocol processing invoked • Multiple interrupts, how to determine who gets service first ─ Priority Ning Weng ECE 526 7 Priorities • • • • • Determining the code which CPU executing at any time Interrupts have different priorities Higher priority code can interrupt lower priority code Kernel software can specify desired interrupt level Possible outcome of priorities processing ─ Livelock: CPU has no efficient power to handle a higher priority load • Packet processing: ─ ─ ─ ─ Device (NIC): receiving packets Protocol stack: processing packets Applications: further processing packet if necessary Question: who should has the highest priority Ning Weng ECE 526 8 Packet Processing Priority • Packet processing priorities: ─ Highest to device driver ─ Lowest to application • Requires queues between interrupt levels ─ Why? ─ What is inside queues? • Processing in higher interrupts should be kept briefly Ning Weng ECE 526 9 Kernel Threads • Thread: piece of software that runs in its own context ─ Similar to process but is light-weight • Different threads can run in different priorities ─ More fine-grained than interrupt levels ─ Scheduling policy allocating CPU to threads • Thread synchronization handles access to shared data ─ Multiple exclusion: only one thread has accesses to data structure ─ Notification: thread blocks until event occurs • How should threads be assigned to layered protocol? ─ ─ ─ ─ One thread per layer One thread per protocol Multiple threads per protocol One thread per packet Ning Weng ECE 526 10 Timer Management • Fundamental function • Timers used for ─ Protocols • ARP for retransmission and cache management • IP for re-assembly • TCP for retransmission ─ Scheduling • Multiple independent timers required ─ Cost can be high Ning Weng ECE 526 11 One Thread per Layer • Different layers have different thread • Allowing priority to be assigned to each layer • A packet is en-queued once per layer • Easy for programmer to understand Ning Weng ECE 526 12 One Thread per Protocol • Each protocol has its own thread and queue • Advantage over one thread per layer ─ Easier for programmer to understand ─ Finer-grain control and each protocol has its own priority Ning Weng ECE 526 13 Multiple Threads per Protocol • Further division of duties • Finer-granularity control and easier for programming • Example division ─ One thread for incoming packets ─ One thread for outgoing packets ─ Thread for management/timing Ning Weng ECE 526 14 One Thread per Packet • Layers introducing overhead ─ Queuing ─ Context switching ─ Requiring many threads • How about packet processed entirely by one thread? ─ “run to completion” programming mode • One thread operation ─ ─ ─ ─ Pre-allocating set of operations Waiting for packet arrive Moving through protocol stack Returning to wait for next packet Ning Weng ECE 526 15 Asynchronous vs. Synchronous • Synchronous programming: thread-structured ─ API using blocking • Main flow-of-control • Explicitly invokes functions as needed ─ API using polling • Function call returns immediately • Performs operations if available • Returns error code otherwise • Asynchronous programming: event-driven ─ Programmer specifying which function to invoke for each event type ─ Programmer has no control over function invocation ─ Difficult to program Ning Weng ECE 526 16 Typical Implementations • Application program ─ Synchronous API using blocking (e.g. socket API) ─ One application thread running while other blocks • Embedded system ─ Synchronous API using polling ─ CPU dedicating to one task • Operating systems ─ Asynchronous API ─ Built on interrupt mechanism Ning Weng ECE 526 17 Summary • Protocol processing implementation on conventional processor ─ Application ─ OS kernel ─ Embedded systems • Packet processing priorities • Two mechanism to control processing priorities ─ Software interrupt ─ Kernel thread abstraction Ning Weng ECE 526 18