Notes on Latency and Payload Graphs

Note the following about the preceding latency and payload graphs:

• Latency is ordered as follows, starting with the lowest latency:
  1. ef_vi using CTPIO
  2. TCPDirect using cut-through CTPIO
     Note: Cut-through CTPIO is supported only for 10 Gb.
  3. TCPDirect using store and forward CTPIO
  4. TCPDirect without CTPIO
  5. Onload with the latency-best profile (which uses cut-through CTPIO)
     Note: Cut-through CTPIO is supported only for 10 Gb.
  6. Onload with the latency profile (which uses store and forward CTPIO, without poisoning).
  7. Kernel.
• The relative latency of CTPIO variants differs by payload:
  • for very small payloads, store and forward CTPIO has similar latency to cut-through CTPIO
    Note: Cut-through CTPIO is supported only for 10 Gb.
  • for intermediate payloads, the latency of store and forward CTPIO gradually increases relative to cut-through CTPIO, tending towards the latency for no CTPIO
  • for large payloads, store and forward CTPIO has similar latency to no CTPIO.
• The step in the graphs for a payload of 200+ bytes occurs when the packet size reaches the PIO threshold of 256 bytes.