SmartNIC Programming Models

Size: px

Start display at page:

Download "SmartNIC Programming Models"

Benedict Fowler
5 years ago
Views:

1 SmartNIC Programming Models Johann Tönsing Open-NFP

2 Agenda SmartNIC hardware Pre-programmed vs. custom (C and/or P4) firmware Programming models / offload models Switching on NIC, with SR-IOV / Virtio data delivery SmartNIC performance + TCO Silicon and datapath software architectures Example code 207 Open-NFP 2

3 Agilio CX SmartNIC Hardware Optimized for standard server-based cloud data centers Low Profile Half Length PCIe form factor, power < 25W Based on Netronome s NFP-4xxx silicon (72 C programmable cores, 8 threads each) 2GB DRAM for lookup tables / state tables (millions of entries) Data plane fully implemented in software 2x 0GbE 2x 25GbE x 40GbE 2x 40GbE Also available: Agilio LX 2x40G / x00g with dual PCIe interfaces, 20 cores, 8GB DRAM 207 Open-NFP 3

4 SmartNIC Firmware: Pre-programmed or Custom SmartNIC with dynamically downloadable firmware 207 Open-NFP 4

5 SmartNIC Firmware: Pre-programmed or Custom OpenStack ONOS ODL Linux BSD vrouter OVS ebpf D P D K OVS / vrouter / ebpf+xdp datapath on host can be accelerated by SmartNIC Firmware / drivers supplied by Netronome SmartNIC with dynamically downloadable firmware 207 Open-NFP 4

6 SmartNIC Firmware: Pre-programmed or Custom OpenStack ONOS ODL Linux BSD vrouter OVS ebpf D P D K OVS / vrouter / ebpf+xdp datapath on host can be accelerated by SmartNIC Firmware / drivers supplied by Netronome SmartNIC with dynamically downloadable firmware Run-Time Firmware can be developed in P4 and/or C Compiler Debugger Custom dataplane app.p4 app.c Editor 207 Open-NFP 4

7 SmartNIC Firmware: Pre-programmed or Custom OpenStack ONOS ODL Linux BSD vrouter OVS ebpf D P D K OVS / vrouter / ebpf+xdp datapath on host can be accelerated by SmartNIC Firmware / drivers supplied by Netronome SmartNIC with dynamically downloadable firmware Run-Time Firmware can be developed in P4 and/or C Compiler Debugger Custom dataplane app.p4 Editor app.c Hybrid - sandbox / plugin concept Example: C plugin embedded in P4 207 Open-NFP 4

8 SmartNIC Firmware: Pre-programmed or Custom OpenStack ONOS ODL Linux BSD vrouter OVS ebpf D P D K OVS / vrouter / ebpf+xdp datapath on host can be accelerated by SmartNIC Firmware / drivers supplied by Netronome SmartNIC with dynamically downloadable firmware Run-Time Firmware can be developed in P4 and/or C Compiler Debugger Custom dataplane 207 Open-NFP app.p4 app.c Editor Integrated Development Environment Edit - Compile - Debug (HW, simulator) Hybrid - sandbox / plugin concept Example: C plugin embedded in P4 4

9 Programming Models 207 Open-NFP 5

10 Programming Models Offload / Acceleration (transparent - drop in replacement) User Mode Kernel Data Path 207 Open-NFP 5

11 Programming Models Offload / Acceleration (transparent - drop in replacement) User Mode Kernel Data Path offload NIC w. NFP Match Data Path Act 207 Open-NFP 5

12 Programming Models Offload / Acceleration (transparent - drop in replacement) User Mode Kernel Data Path offload NIC w. NFP Match Data Path Act 207 Open-NFP 5

13 Programming Models Offload / Acceleration (transparent - drop in replacement) User Mode VM / App Kernel Data Path offload NIC w. NFP Match Data Path Act 207 Open-NFP 5

14 Programming Models Offload / Acceleration (transparent - drop in replacement) User Mode VM / App Kernel Data Path NIC w. NFP offload fallback Match Data Path Act 207 Open-NFP 5

15 Programming Models Offload / Acceleration (transparent - drop in replacement) User Mode VM / App Kernel Data Path NIC w. NFP offload fallback Match Data Path Act t OVS/OpenFlow Match, Action, Tunnels Contrail vrouter Match, Action, Tunnels Conntrack (Firewall) ebpf / XDP Open-NFP 5

16 Programming Models Offload / Acceleration (transparent - drop in replacement) Calling APIs Compatible / Open Sourced vs. Vendor Extension User Mode VM / App User Mode App API Kernel Data Path Kernel Data Path NIC w. NFP offload fallback NIC w. NFP Match Data Path Act Match Data Path Act t OVS/OpenFlow Match, Action, Tunnels Contrail vrouter Match, Action, Tunnels Conntrack (Firewall) ebpf / XDP Open-NFP 5

17 Programming Models Offload / Acceleration (transparent - drop in replacement) Calling APIs Compatible / Open Sourced vs. Vendor Extension User Mode VM / App User Mode App API Kernel Data Path Kernel Data Path NIC w. NFP offload fallback NIC w. NFP Match Data Path Act Match Data Path Act t OVS/OpenFlow Match, Action, Tunnels Contrail vrouter Match, Action, Tunnels Conntrack (Firewall) ebpf / XDP... t DPDK Poll Mode Driver ebpf / XDP APIs Flow APIs - Match / Act / Tunnel Load Balancing APIs Crypto APIs Open-NFP 5

18 Programming Models Offload / Acceleration (transparent - drop in replacement) Calling APIs Compatible / Open Sourced vs. Vendor Extension User Mode Kernel VM / App Data Path User Mode Kernel App API Data Path New OPNFV! NIC w. NFP offload fallback NIC w. NFP Data Path Data Path Match Act Match Act t OVS/OpenFlow Match, Action, Tunnels Contrail vrouter Match, Action, Tunnels t DPDK Poll Mode Driver ebpf / XDP APIs Flow APIs - Match / Act / Tunnel Conntrack (Firewall) ebpf / XDP... Load Balancing APIs Crypto APIs Open-NFP 5

19 Programming Models Offload / Acceleration (transparent - drop in replacement) Calling APIs Compatible / Open Sourced vs. Vendor Extension Flexible Datapath Abstraction OpenFlow 2.x, P4, PIF, ebpf... User Mode Kernel VM / App Data Path User Mode Kernel App API Data Path New OPNFV! User Mode Kernel App API Data Path offload fallback NIC w. NFP NIC w. NFP NIC w. NFP Data Path Data Path Data Path Parse Match Match State Match Act Match Act Act t OVS/OpenFlow Match, Action, Tunnels Contrail vrouter Match, Action, Tunnels Conntrack (Firewall) ebpf / XDP... t DPDK Poll Mode Driver ebpf / XDP APIs Flow APIs - Match / Act / Tunnel Load Balancing APIs Crypto APIs... t Protocol agnostic flexible parsing Arbitrary arrangement of matching tables Matching without tables State storage / retrieval Complex actions Event handling 207 Open-NFP 5

20 Programming Models Offload / Acceleration (transparent - drop in replacement) Calling APIs Compatible / Open Sourced vs. Vendor Extension Flexible Datapath Abstraction OpenFlow 2.x, P4, PIF, ebpf... Hybrid: Datapath Extensions in / NFP In C, P4 / PIF,... User Mode Kernel VM / App Data Path User Mode Kernel App API Data Path New OPNFV! User Mode Kernel App API Data Path User Mode Kernel App API Data Path offload fallback NIC w. NFP NIC w. NFP NIC w. NFP NIC w. NFP Data Path Data Path Data Path Parse Match Match State Data Path App Match Act Match Act Act Match Act t OVS/OpenFlow Match, Action, Tunnels Contrail vrouter Match, Action, Tunnels Conntrack (Firewall) ebpf / XDP... t DPDK Poll Mode Driver ebpf / XDP APIs Flow APIs - Match / Act / Tunnel Load Balancing APIs Crypto APIs... t Protocol agnostic flexible parsing Arbitrary arrangement of matching tables Matching without tables State storage / retrieval Complex actions Event handling t Custom tunnel Custom action Custom matching Open-NFP 5

21 Traditional Model: SR-IOV x86 Userspace (Nova, Neutron) SR-IOV Configuration Configuration from cloud management system SR-IOV PCIe Traditional NIC OVS Kernel DP Match/Act SR-IOV Traffic Director 207 Open-NFP 6

22 Traditional Model: SR-IOV x86 Userspace (Nova, Neutron) SR-IOV Configuration 2 Configuration from cloud management system SR-IOV 2 Hardware configuration PCIe Traditional NIC OVS Kernel DP Match/Act SR-IOV Traffic Director 207 Open-NFP 6

23 Traditional Model: SR-IOV x86 Userspace (Nova, Neutron) SR-IOV Configuration 2 Configuration from cloud management system SR-IOV 2 Hardware configuration PCIe Traditional NIC OVS Kernel DP Match/Act SR-IOV Traffic Director 207 Open-NFP 6

Traditional Model: SR-IOV x86 Userspace (Nova, Neutron) SR-IOV Configuration 2 Configuration from cloud management system SR-IOV 2 Hardware configuration PCIe Low

24 Traditional Model: SR-IOV x86 Userspace (Nova, Neutron) SR-IOV Configuration 2 Configuration from cloud management system SR-IOV 2 Hardware configuration PCIe Low expressiveness (MAC/VLAN based traffic directing) High performance Poor manageability (no VM migration) Traditional NIC OVS Kernel DP Match/Act SR-IOV Traffic Director 207 Open-NFP 6

25 Traditional Model: Host Runs Datapath (Nova, Neutron) x86 Userspace Datapath P4, Open vswitch, vrouter Control Agent CLI or equivalent x86 Kernel Matching Execute Action Host - guest channel virtio or other Configuration via network protocol or CLI PCIe Traditional NIC OVS Kernel DP Match/Act 207 Open-NFP 7

26 Traditional Model: Host Runs Datapath (Nova, Neutron) or equivalent x86 Userspace x86 Kernel Datapath P4, Open vswitch, vrouter Matching 2 Control Agent CLI Execute Action Host - guest channel virtio or other Configuration via network protocol or CLI PCIe 2 Userspace agent populates table entries Traditional NIC OVS Kernel DP Match/Act 207 Open-NFP 7

27 Traditional Model: Host Runs Datapath (Nova, Neutron) or equivalent x86 Userspace x86 Kernel Datapath P4, Open vswitch, vrouter Matching 2 Control Agent CLI Execute Action Host - guest channel virtio or other Configuration via network protocol or CLI PCIe 2 Userspace agent populates table entries Traditional NIC OVS Kernel DP Match/Act 207 Open-NFP 7

28 Traditional Model: Host Runs Datapath (Nova, Neutron) or equivalent x86 Userspace x86 Kernel Datapath P4, Open vswitch, vrouter Matching 2 Control Agent CLI Execute Action Host - guest channel virtio or other Configuration via network protocol or CLI PCIe 2 Userspace agent populates table entries Traditional NIC High expressiveness (software) Consumes many cores and/or exhibits low performance OVS Kernel DP Match/Act 207 Open-NFP 7

29 Offload Model: Agilio OVS Acceleration x86 Userspace (Nova, Neutron) Open vswitch Subsystem OVS CLI Callable API OVS Agent OpenFlow x86 Kernel Configuration via controller, CLI, or Callable API OVS Kernel DP Match/Act Execute Action SR-IOV / virtio VFs SR-IOV VFs PCIe Agilio SmartNIC OVS Kernel DP Match/Act 207 Open-NFP 8

30 Offload Model: Agilio OVS Acceleration x86 Userspace (Nova, Neutron) Open vswitch Subsystem OVS CLI Callable API OVS Agent OpenFlow x86 Kernel Configuration via controller, CLI, or Callable API OVS Kernel DP Match/Act Execute Action SR-IOV / virtio VFs SR-IOV VFs PCIe Agilio SmartNIC Open vswitch Datapath OVS Kernel DP Match/Act Execute Action (e.g. Entunnel, Deliver to VM, Send to Port) 207 Open-NFP 8

31 Offload Model: Agilio OVS Acceleration x86 Userspace (Nova, Neutron) Open vswitch Subsystem OVS CLI Callable API OVS Agent OpenFlow x86 Kernel Configuration via controller, CLI, or Callable API OVS Kernel DP Match/Act Execute Action SR-IOV / virtio VFs SR-IOV VFs Agilio SmartNIC Miss PCIe Open vswitch Datapath OVS Kernel DP Match/Act Execute Action (e.g. Entunnel, Deliver to VM, Send to Port) 207 Open-NFP 8

32 Offload Model: Agilio OVS Acceleration x86 Userspace (Nova, Neutron) Open vswitch Subsystem OVS CLI Callable API OVS Agent OpenFlow Miss x86 Kernel Configuration via controller, CLI, or Callable API OVS Kernel DP Match/Act Execute Action SR-IOV / virtio VFs SR-IOV VFs Agilio SmartNIC Miss PCIe Open vswitch Datapath OVS Kernel DP Match/Act Execute Action (e.g. Entunnel, Deliver to VM, Send to Port) 207 Open-NFP 8

33 Offload Model: Agilio OVS Acceleration x86 Userspace (Nova, Neutron) Open vswitch Subsystem OVS CLI Callable API OVS Agent OpenFlow 2 Miss x86 Kernel Configuration via controller, CLI, or Callable API OVS Kernel DP Match/Act Execute Action SR-IOV / virtio VFs SR-IOV VFs 2 OVS userspace agent populates kernel cache Agilio SmartNIC Miss PCIe Open vswitch Datapath OVS Kernel DP Match/Act Execute Action (e.g. Entunnel, Deliver to VM, Send to Port) 207 Open-NFP 8

34 Offload Model: Agilio OVS Acceleration x86 Userspace (Nova, Neutron) Open vswitch Subsystem OVS CLI Callable API OVS Agent OpenFlow 2 Miss x86 Kernel Configuration via controller, CLI, or Callable API OVS Kernel DP Match/Act Execute Action SR-IOV / virtio VFs SR-IOV VFs 2 OVS userspace agent populates kernel cache 3 Offload datapath: copy match tables, sync stats Agilio SmartNIC 3 Miss PCIe Open vswitch Datapath OVS Kernel DP Match/Act Execute Action (e.g. Entunnel, Deliver to VM, Send to Port) 207 Open-NFP 8

35 Offload Model: Agilio OVS Acceleration x86 Userspace (Nova, Neutron) Open vswitch Subsystem OVS CLI Callable API OVS Agent OpenFlow 2 Miss x86 Kernel Configuration via controller, CLI, or Callable API OVS Kernel DP Match/Act Execute Action SR-IOV / virtio VFs SR-IOV VFs 2 OVS userspace agent populates kernel cache 3 Offload datapath: copy match tables, sync stats Agilio SmartNIC 3 Miss PCIe Open vswitch Datapath 4 Flow tracking: per-microflow state learning Self Learning Exact Match Flow Tracker Miss OVS Kernel DP Match/Act Execute Action (e.g. Entunnel, Deliver to VM, Send to Port) 4 Hit 207 Open-NFP 8

36 Offload Model: Agilio OVS Acceleration x86 Userspace (Nova, Neutron) Open vswitch Subsystem OVS CLI Callable API OVS Agent OpenFlow 2 Miss x86 Kernel Configuration via controller, CLI, or Callable API OVS Kernel DP Match/Act Conn track FTP SIP Execute Action SR-IOV / virtio VFs SR-IOV VFs 2 OVS userspace agent populates kernel cache 3 Offload datapath: copy match tables, sync stats Agilio SmartNIC 3 Miss PCIe Open vswitch Datapath 4 Flow tracking: per-microflow state learning Self Learning Exact Match Flow Tracker Miss OVS Kernel DP Match/Act Execute Action (e.g. Entunnel, Deliver to VM, Send to Port) 4 Hit 207 Open-NFP 8

37 Offload Model: Agilio OVS Acceleration x86 Userspace (Nova, Neutron) Open vswitch Subsystem OVS CLI Callable API OVS Agent OpenFlow 2 Miss x86 Kernel Configuration via controller, CLI, or Callable API OVS Kernel DP Match/Act Conn track FTP SIP Execute Action SR-IOV / virtio VFs SR-IOV VFs 2 OVS userspace agent populates kernel cache 3 Offload datapath: copy match tables, sync stats Agilio SmartNIC 3 Miss PCIe Open vswitch Datapath 4 Flow tracking: per-microflow state learning Self Learning Exact Match Flow Tracker Miss OVS Kernel DP Match/Act Conn track Execute Action (e.g. Entunnel, Deliver to VM, Send to Port) 4 Hit 207 Open-NFP 8

38 Offload Model: Agilio OVS Acceleration x86 Userspace (Nova, Neutron) Open vswitch Subsystem OVS CLI Callable API OVS Agent OpenFlow 2 Miss x86 Kernel Configuration via controller, CLI, or Callable API OVS Kernel DP Match/Act Conn track FTP SIP Execute Action SR-IOV / virtio VFs SR-IOV VFs 2 OVS userspace agent populates kernel cache 3 Offload datapath: copy match tables, sync stats Agilio SmartNIC 3 Miss 5 PCIe Open vswitch Datapath 4 Flow tracking: per-microflow state learning 5 Offload connection tracking: synchronize state Self Learning Exact Match Flow Tracker Miss OVS Kernel DP Match/Act Conn track Execute Action (e.g. Entunnel, Deliver to VM, Send to Port) 4 Hit 207 Open-NFP 8

39 Offload Model: Agilio OVS Acceleration x86 Userspace (Nova, Neutron) Open vswitch Subsystem OVS CLI Callable API OVS Agent OpenFlow 2 Miss x86 Kernel DP Ext. Configuration via controller, CLI, or Callable API OVS Kernel DP Match/Act Conn track FTP SIP Execute Action 6 SR-IOV / virtio VFs SR-IOV VFs 2 OVS userspace agent populates kernel cache 3 Offload datapath: copy match tables, sync stats 4 Flow tracking: per-microflow state learning 5 Offload connection tracking: synchronize state Agilio SmartNIC Self Learning Exact Match Flow Tracker 3 Miss Miss OVS Kernel DP Match/Act 5 Conn track Open vswitch Datapath Execute Action (e.g. Entunnel, Deliver to VM, Send to Port) PCIe 6 Datapath Extension or Plugin P4 / C in Sandbox 6 Datapath extension software 4 Hit 207 Open-NFP 8

40 Offload Model: Agilio OVS Acceleration x86 Userspace (Nova, Neutron) Open vswitch Subsystem OVS CLI Callable API OVS Agent OpenFlow 2 Miss x86 Kernel DP Ext. Configuration via controller, CLI, or Callable API OVS Kernel DP Match/Act Conn track FTP SIP Execute Action 6 SR-IOV / virtio VFs SR-IOV VFs 2 OVS userspace agent populates kernel cache 3 Offload datapath: copy match tables, sync stats 4 Flow tracking: per-microflow state learning 5 Offload connection tracking: synchronize state Agilio SmartNIC Self Learning Exact Match Flow Tracker 3 Miss Miss OVS Kernel DP Match/Act 5 Conn track Open vswitch Datapath Execute Action (e.g. Entunnel, Deliver to VM, Send to Port) PCIe 6 Datapath Extension or Plugin P4 / C in Sandbox 6 Datapath extension software 4 Hit 207 Open-NFP 8

41 Offload Model: Agilio OVS Acceleration x86 Userspace (Nova, Neutron) Open vswitch Subsystem OVS CLI Callable API OVS Agent OpenFlow 2 Miss x86 Kernel DP Ext. Best of all worlds Configuration via controller, CLI, or Callable API - Performance of SR-IOV OVS Kernel DP Match/Act Conn track FTP SIP Execute Action 6 SR-IOV / virtio VFs SR-IOV VFs - Flexibility of Virtio (VM migration) 2 OVS userspace agent populates kernel cache - Performance and core saving of switching on SmartNIC 3 Offload datapath: copy match tables, sync stats 4 Flow tracking: per-microflow state learning 5 Offload connection tracking: synchronize state Agilio SmartNIC Self Learning Exact Match Flow Tracker 3 Miss Miss OVS Kernel DP Match/Act 5 Conn track Open vswitch Datapath Execute Action (e.g. Entunnel, Deliver to VM, Send to Port) PCIe 6 Datapath Extension or Plugin P4 / C in Sandbox 6 Datapath extension software 4 Hit 207 Open-NFP 8

42 Example: Throughput vs. Number of Rules Millions of Packets per Second Wildcard Rules 000 Wildcard Rules 0000 Wildcard Rules Wildcard Rules L2/L3 Forwarding to 8 VMs with 64K Flows 2 Cores 2 Cores 5X Throughput Improvement + 90% Savings Core 5 OVS in Kernel Space OVS in User Space on DPDK OVS Offloaded to Agilio CX 207 Open-NFP 9

43 Example: Throughput vs. Number of Rules Millions of Packets per Second Wildcard Rules 000 Wildcard Rules 0000 Wildcard Rules Wildcard Rules L2/L3 Forwarding to 8 VMs with 64K Flows 2 Cores 2 Cores 5X Throughput Improvement + 90% Savings Core 5 OVS in Kernel Space OVS in User Space on DPDK OVS Offloaded to Agilio CX 207 Open-NFP 9

44 Tested Scenario: Server Core Allocation Unaccelerated OVS (Kernel / User Mode) Agilio OVS OVS 207 Open-NFP 0

45 Tested Scenario: Server Core Allocation Unaccelerated OVS (Kernel / User Mode) Agilio OVS OVS Typical result: replace 3-6 racks with rack! 207 Open-NFP 0

46 Tested Scenario: Server Core Allocation Unaccelerated OVS (Kernel / User Mode) Agilio OVS OVS Benefits for your use case: Typical result: replace 3-6 racks with rack! 207 Open-NFP 0

47 P4 Datapath on SmartNIC (Nova, Neutron) x86 Server P4 Control CLI P4 Control Agent or successor SR-IOV / virtio VFs SR-IOV VFs PCIe Agilio SmartNIC P4 Matching 207 Open-NFP

48 P4 Datapath on SmartNIC (Nova, Neutron) x86 Server P4 Control CLI P4 Control Agent or successor Configuration via control protocol or CLI SR-IOV / virtio VFs SR-IOV VFs PCIe Agilio SmartNIC P4 Matching 207 Open-NFP

49 P4 Datapath on SmartNIC (Nova, Neutron) x86 Server P4 Control CLI P4 Control Agent or successor Configuration via control protocol or CLI SR-IOV / virtio VFs SR-IOV VFs PCIe Agilio SmartNIC P4 Datapath P4 Matching Execute Action (e.g. Entunnel, Deliver to VM, Send to Port) 207 Open-NFP

50 P4 Datapath on SmartNIC (Nova, Neutron) x86 Server P4 Control CLI P4 Control Agent or successor Configuration via control protocol or CLI 2 SR-IOV / virtio VFs SR-IOV VFs 2 Agent populates tables in SmartNIC datapath PCIe Agilio SmartNIC P4 Datapath P4 Matching Execute Action (e.g. Entunnel, Deliver to VM, Send to Port) 207 Open-NFP

51 Traditional vs. SmartNIC Accelerated Networking Forwarding / Virtual Switching Technology SR-IOV Intelligent Datapath P4 / vrouter / Open vswitch 207 Open-NFP 2

52 Traditional vs. SmartNIC Accelerated Networking Forwarding / Virtual Switching Technology SR-IOV Intelligent Datapath P4 / vrouter / Open vswitch Traditional Approach (Unaccelerated) Limited expressiveness to direct traffic to VMs (no support for general match/action rules, tunnel termination, stateful firewalling) High throughput No VM migration support (difficult to manage) High expressiveness - match/action, tunnels, stateless/stateful firewalling etc. Limited throughput High utilization (e.g. 50% of cores) 207 Open-NFP 2

53 Traditional vs. SmartNIC Accelerated Networking Forwarding / Virtual Switching Technology SR-IOV Intelligent Datapath P4 / vrouter / Open vswitch Traditional Approach (Unaccelerated) Limited expressiveness to direct traffic to VMs (no support for general match/action rules, tunnel termination, stateful firewalling) High throughput No VM migration support (difficult to manage) High expressiveness - match/action, tunnels, stateless/stateful firewalling etc. Limited throughput High utilization (e.g. 50% of cores) Fully Programmable SmartNIC Accelerated Approach High expressiveness - match/action, tunnels, stateless/stateful firewalling etc. and SR-IOV based data delivery to VMs High throughput Virtio supporting VM migration (facilitating cloud optimization and upgrading) Higher throughput (~5x higher) Lower utilization (~0x lower) 207 Open-NFP 2

54 Flow Processor Silicon Architecture Network Flow Processor 4xxx (used on Agilio-CX SmartNICs) Highly parallel multithreaded architecture (8 threads / core) for high throughput Purpose built Flow Processing Cores (72) maximize flexibility H/W accelerators further maximize efficiency (throughput/watt) Fully software defined feature set - examples: PCIe Gen3 x8 Network and PCIe SR-IOV / virtio RX/TX with stateless offloads Flexible tunneling support (e.g. VXLAN, GRE, VLAN, MPLS, NSH) Flexible Match/Action processing - many packet fields / protocols Highly scalable and fine grained security policies External DDR3 accommodates millions of flows / rules Up to 8GB DRAM Convenient programmability using P4 and C Ethernet 0G, 40G, 207 Open-NFP 3

55 Flow Processor Silicon Architecture Network Flow Processor 4xxx (used on Agilio-CX SmartNICs) Highly parallel multithreaded architecture (8 threads / core) for high throughput Purpose built Flow Processing Cores (72) maximize flexibility H/W accelerators further maximize efficiency (throughput/watt) Fully software defined feature set - examples: PCIe Gen3 x8 Network and PCIe SR-IOV / virtio RX/TX with stateless offloads Flexible tunneling support (e.g. VXLAN, GRE, VLAN, MPLS, NSH) Flexible Match/Action processing - many packet fields / protocols Highly scalable and fine grained security policies External DDR3 accommodates millions of flows / rules Up to 8GB DRAM Convenient programmability using P4 and C Ethernet 0G, 40G, 207 Open-NFP 3

Flow Processor Silicon Architecture Network Flow Processor 4xxx (used on Agilio-CX SmartNICs) Highly parallel multithreaded architecture (8 threads / core) for high throughput Purpose built Flow

56 Flow Processor Silicon Architecture Network Flow Processor 4xxx (used on Agilio-CX SmartNICs) Highly parallel multithreaded architecture (8 threads / core) for high throughput Purpose built Flow Processing Cores (72) maximize flexibility H/W accelerators further maximize efficiency (throughput/watt) Fully software defined feature set - examples: PCIe Gen3 x8 Network and PCIe SR-IOV / virtio RX/TX with stateless offloads Flexible tunneling support (e.g. VXLAN, GRE, VLAN, MPLS, NSH) Flexible Match/Action processing - many packet fields / protocols Highly scalable and fine grained security policies External DDR3 accommodates millions of flows / rules Up to 8GB DRAM Convenient programmability using P4 and C Ethernet 0G, 40G, 207 Open-NFP 3

57 SmartNIC Datapath Worker Software Architecture Load balancer distributes each packet to next available thread for optimum throughput Hardware assisted reordering ensures packet order is maintained Flow tracker statefully learns / tracks millions of sessions Matching performed using DRAM-backed tables - capacity millions of entries Actions efficiently performed in on-chip memory = Ring / Work Queue (multi producer / consumer) net or PCIE Load Balance Parse Flow Tracker Learn microflows Match Act Cache action P4 Datapath Run to completion Re- order net or PCIE Pool of worker threads on flow processing cores 207 Open-NFP 4

58 SmartNIC Datapath Worker Software Architecture Load balancer distributes each packet to next available thread for optimum throughput Hardware assisted reordering ensures packet order is maintained Flow tracker statefully learns / tracks millions of sessions Matching performed using DRAM-backed tables - capacity millions of entries Actions efficiently performed in on-chip memory = Ring / Work Queue (multi producer / consumer) net or PCIE Load Balance Parse Flow Tracker Learn microflows Match Act Cache action P4 Datapath Run to completion C Plugin Re- order net or PCIE Pool of worker threads on flow processing cores 207 Open-NFP 4

59 SmartNIC Datapath Worker Software Architecture Load balancer distributes each packet to next available thread for optimum throughput Hardware assisted reordering ensures packet order is maintained Flow tracker statefully learns / tracks millions of sessions Matching performed using DRAM-backed tables - capacity millions of entries Actions efficiently performed in on-chip memory = Ring / Work Queue (multi producer / consumer) net or PCIE Load Balance Parse Flow Tracker Learn microflows Match Act Cache action C XX P4 Datapath Run to completion C Plugin Re- order net or PCIE Pool of worker threads on flow processing cores 207 Open-NFP 4

60 Datapath Distributed over Microengines Worker uses ring to forward packets to other microengine(s) = Ring / Work Queue (multi producer / consumer) net or PCIE Load Balance Parse C / P4 Datapath Run to completion Flow Tracker Learn microflows Cache action Match Act Re- order net or PCIE Pool of worker threads on microengines 207 Open-NFP 5

61 Datapath Distributed over Microengines Worker uses ring to forward packets to other microengine(s) Coprocessor microengines = Ring / Work Queue (multi producer / consumer) C Code net or PCIE Load Balance Parse C / P4 Datapath Run to completion Flow Tracker Learn microflows Cache action Match Act Re- order net or PCIE Pool of worker threads on microengines 207 Open-NFP 5

62 Datapath Distributed over Microengines Worker uses ring to forward packets to other microengine(s) Coprocessor microengines = Ring / Work Queue (multi producer / consumer) C Code net or PCIE Load Balance Parse C / P4 Datapath Run to completion Flow Tracker Learn microflows Cache action Match Act Re- order net or PCIE Pool of worker threads on microengines 207 Open-NFP 5

63 Datapath Distributed over Microengines Worker uses ring to forward packets to other microengine(s) Coprocessor microengines = Ring / Work Queue (multi producer / consumer) C Code net or PCIE Load Balance Parse C / P4 Datapath Run to completion Flow Tracker Learn microflows Cache action Match Act Re- order net or PCIE Pool of worker threads on microengines 207 Open-NFP 5

64 Datapath Distributed over Microengines Worker uses ring to forward packets to other microengine(s) = Ring / Work Queue (multi producer / consumer) Coprocessor microengines C Code Coprocessor microengines C Code net or PCIE Load Balance Parse C / P4 Datapath Run to completion Flow Tracker Learn microflows Cache action Match Act Re- order net or PCIE Pool of worker threads on microengines 207 Open-NFP 5

65 Datapath Distributed over Microengines Worker uses ring to forward packets to other microengine(s) Coprocessor microengines Coprocessor microengines = Ring / Work Queue (multi producer / consumer) C Code C Code Crypto Crypto net or PCIE Load Balance Parse C / P4 Datapath Run to completion Flow Tracker Learn microflows Cache action Match Act Re- order net or PCIE Pool of worker threads on microengines 207 Open-NFP 5

66 Example: P4 main implementing a simple NIC header_type eth_hdr { fields { dst : 48; src : 48; etype : 6; } } header eth_hdr eth; parser start { return eth_parse; } parser eth_parse { extract(eth); return ingress; } action drop_act() { drop(); } action fwd_act(port) { modify_field(standard_metadata.egress_spec, port); } table in_tbl { reads { standard_metadata.ingress_port : exact; } actions { fwd_act; drop_act; } } control ingress { apply(in_tbl); } 207 Open-NFP 6

67 Example: C Code 207 Open-NFP 7

68 Example of Fully Customized Datapath (P4 / C) Control App Populate tables, display statistics P4 / C Development Environment Edit - Debug Server (x86 - Linux) Run-Time Interface Agent VNF Kernel Mode (C) netdev 2 VNF User Mode (C) DPDK Concepts: P4 and C running on SmartNIC implements datapath - e.g. defines protocols, match / action behavior Datapath steers traffic to VNFs running on x86 server and on SmartNIC Agilio SmartNIC Match Protocol RT I/F Helper Other Security µvnf (C) Timestamp µvnf Latency Stats µvnf (C) Timestamp µvnf Latency Stats µvnf (C) PCIe TCP Meter P4 Datapath with C Run to completion 207 Open-NFP 8

69 Next Steps Use Agilio SmartNICs with existing dataplanes Use Agilio OVS (with / without Conntrack) Use Agilio Contrail vrouter Use Agilio ebpf/xdp Program Agilio SmartNICs (following sessions - SDK, P4, OVS, ebpf/xdp) Program using P4, C, ebpf/xdp Participate in open source and standards evolution: openstack.org, openvswitch.org, opencontrail.org, p4.org, iovisor.org, open-nfp.org, opennetworking.org, opensourcesdn.org, opnfv.org, linuxfoundation.org Examples: VNF offload API (@ OPNFV + other bodies) Increase flexibility, improve performance, free up server resources! 207 Open-NFP 9

70 Thank You! More information: open-nfp.org 207 Open-NFP 20

SmartNIC Programming Models

SmartNIC Programming Models Johann Tönsing 206--09 206 Open-NFP Agenda SmartNIC hardware Pre-programmed vs. custom (C and/or P4) firmware Programming models / offload models Switching on NIC, with SR-IOV