Montek Singh COMP Nov 10, 2011

 Design questions at various leves ◦ Network Adapter design ◦ Network level: topology and routing ◦ Link level: synchronization and timing  Discussion ◦ benefits and challenges ◦ key research problems

 Analogous to the network adapters for Wifi/Ethernet

 Sockets abstraction ◦ orthogonalize computation and communication  hide networking details  the core-side interface provides send/receive commands  Standardization ◦ Open Core Protocols (OCP) ◦ Virtual Component Interface (VCI) ◦ Advanced eXtensible Interface (AXI) ◦ Device Transaction Level (DTL)

 Adapter responsibilities ◦ encapsulation of traffic  for the underlying communication ◦ management of services  that the network provides  Implementations: several ◦ Muttersbach et al. ◦ Bjerregaard et al. ◦ Radulescu et al. ◦ HERMES ◦ Bhojwani/Mahapatra ◦ …

 Network responsibility ◦ deliver messages from source to destination ◦ hardware support for basic communication commands (send/receive) ◦ well-built network should appear as a logical wire [Dally/Towles 2001]  Two main considerations ◦ topology  layout and connectivity ◦ protocol  how nodes and links are used, routing etc.

 Network responsibility ◦ deliver messages from source to destination ◦ hardware support for basic communication commands (send/receive) ◦ well-built network should appear as a logical wire [Dally/Towles 2001]  Two main considerations ◦ topology  layout and connectivity ◦ protocol  how nodes and links are used, routing etc.

 thin (spread) layout vs. concentrated

 Circuit vs. packet switching ◦ circuit: the entire path is set up and reserved for the entire duration of data transport ◦ packet: each packet is forwarded on a per-hop basis  Connection-oriented vs. connectionless ◦ connection: dedicated logical path established prior to data transport  may or may not be circuit-switched (logically)

 Deterministic vs. adaptive routing ◦ deterministic: path is determined by source and destination pair, alone ◦ adaptive: dynamically determined, incl. arbitration, congestion, load balancing, etc.  Minimal vs. nonminimal routing ◦ always shortest path or not?  Delay vs. loss model ◦ delay model: data packets never dropped, but may be delayed ◦ loss model: data packets may be dropped due to congestion, requiring retransmission

 Central vs. distributed control ◦ central: global routing decisions (e.g., bus control) ◦ distributed: decisions made locally

 Flow control: control the flow of data with some objective (delay, loss, etc.)  Virtual channels: 2 to 16 VCs per physical channel! ◦ avoid deadlocks since they are mutually independent ◦ optimize wire utilization by letting several VCs share a wire ◦ improve performance because of fewer stalls ◦ provide differentiated services for QoS

 Synchronization ◦ different clock domains ◦ maybe network is asynchronous  Challenges: ◦ metastability ◦ arbitration ◦ synchronization

 Globally Asynchronous Locally Synchronous ◦ network is elastic, asynchronous ◦ cores may be clocked

 Why buffering?  How much buffering?

 Other than electrical?  How about optical?

 Other than electrical?  How about optical?

 Delay-insensitive encoding? ◦ very robust  Energy-efficient encoding? ◦ save energy  Handshaking? ◦ two-phase vs. four-phase