unit1.ppt
- 1. CS 6290 Many-core & Interconnect Milos Prvulovic Fall 2007
- 2. Interconnection Networks • Classification: Shared Medium or Switched
- 3. Shared Media Networks • Need arbitration to decide who gets to talk • Arbitration can be centralized or distributed • Centralized not used much for networks – Special arbiter device (or must elect arbiter) – Good performance if arbiter far away? Nah. • Distributed arbitration – Check if media already used (carrier sensing) – If media not used now, start sending – Check if another also sending (collision detection) – If collision, wait for a while and retry • “For a while” is random (otherwise collisions repeat forever) • Exponential back-off to avoid wasting bandwidth on collisions
- 4. Switched Networks • Need switches – Introduces switching overheads • No time wasted on arbitration and collisions • Multiple transfers can be in progress – If they use different links, of course • Circuit or Packet Switching – Circuit switching: end-to-end connections • Reserves links for a connection (e.g. phone network) – Packet switching: each packet routed separately • Links used only when data transferred (e.g. Internet Protocol)
- 5. Routing • Shared media has trivial routing (broadcast) • In switched media we can have – Source-based (source specifies route) – Virtual circuits (end-to-end route created) • When connection made, set up route • Switches forward packets along the route – Destination-based (source specifies destination) • Switches must route packet toward destination • Also can be classified into – Deterministic (one route from a source to a destination) – Adaptive (different routes can be used)
- 6. Routing Methods for Switches • Store-and-Forward – Switch receives entire packet, then forwards it – If error occurs when forwarding, switch can re-send • Wormhole routing – Packet consists of flits (a few bytes each) – First flit contains header w/ destination address – Switch gets header, decides where to forward – Other flits forwarded as they arrive – Looks like packet worming through network – If an error occurs along the way, sender must re-send • No switch has the entire packet to re-send it
- 7. Cut-Through Routing • What happens when link busy? – Header arrives to switch, but outgoing link busy – What do we do with the other flits of the packet? • Wormhole routing: stop the tail when head stops – Now each flit along the way blocks the a link – One busy link creates other busy links => traffic jam • Cut-Through Routing – If outgoing link busy, receive and buffer incoming flits – The buffered flits stay there until link becomes free – When link free, the flits start worming out of the switch – Need packet-sized buffer space in each switch • Wormhole Routing switch needs to buffer only one flit
- 8. Routing: Network Latency • Switch Delay – Time from incoming to outgoing link in a switch • Switches – Number of switches along the way • Transfer time – Time to send the packet through a link • Store-and-Forward end-to-end transfer time – (Switches*SwitchDelay)+(TransferTime*(Switches+1)) • Wormhole or Cut-Through end-to-end transfer time – (Switches*SwitchDelay) + TransferTime – Much better if there are many switches along the way – See the example on page 811
- 9. Switch Technology • What do we want in a switch – Many input and output links • Usually number of input and output links the same – Low contention inside the switch • Best if there is none (only external links cause contention) – Short switching delay • Crossbar – Very low switching delay, no internal contention – Complexity grows as square of number of links • Can not have too many links (e.g. up to 64 in and 64 out)
- 10. Switch Technology • What do we want in a switch – Many input and output links • Usually number of input and output links the same – Low contention inside the switch • Best if there is none (only external links cause contention) – Short switching delay • Crossbar – Very low switching delay, no internal contention – Complexity grows as square of number of links • Can not have too many links (e.g. up to 64 in and 64 out) • Omega Network – Build switches with more ports using small crossbars – Lower complexity per link, but longer delay and more contention
- 12. Network Topology
- 13. Network Topology • What do we want in a network topology – Many nodes, high bandwidth, low contention, low latency – Low latency: few switches along any route • For each (src, dest) pair, we choose shortest route • Longest such route over all (src,dst) pairs: network diameter • We want networks with small diameter! – Low contention: high aggregate bandwidth • Divide network into two groups, each with half the nodes • Total bandwidth between groups is bisection bandwidth • Actually, we use the minimum over all such bisections
- 14. On-Chip Networks • We’ll have many cores on-chip – Need switched network to provide bandwidth • Need to map well onto chip surface – E.g. hypercube is not great • Mesh or grid should work well, torus OK too – Limited ports per switch (CPU & 4 neighbors) – All links short (going to neighbors) – Many parallel algorithms map well onto grids • Matrices, grids, etc.
- 15. Trouble with shared caches • Private caches OK – Each placed with its own processor • We want a shared cache, too – Fits more data than if broken into private caches • Private caches replicate data – Dynamically shared • Threads that need more space get more space • But how do we make a shared cache fast
- 16. Trouble with shared caches • Private caches OK – Each placed with its own processor • We want a shared cache, too – Fits more data than if broken into private caches • Private caches replicate data – Dynamically shared • Threads that need more space get more space • But how do we make a shared cache fast
- 17. Non-uniform Cache Arch. (NUCA) Bank Switch CPU Request 0x….3 Request 0x….C
- 18. S-NUCA Perormance • Fast access to nearby banks • Slow access to far-away banks • Average better than worst-case
- 20. D-NUCA Perormance • Fast access to nearby banks • Slow access to far-away banks • Average much better than worst-case • But we keep moving blocks – Lots of power-hungry activity • Need smart policies for block migration – Move blocks less frequently – But get most of the benefit of being able to move
- 21. D-NUCA Issues • Blocks keep moving, how do we find them? • One solution: Use an on-chip directory! – Use direct mapping to assign a home bank – If we don’t know where the block is, ask the home bank – If we move the block, tell the home bank – If we think we know where the block is, look there. If it’s been moved, ask home bank