1. Data Center Benchmarking Drafts Lucien Avramov, Cisco Jacob Rapp, Cisco July 2013 IETF 87 - Berlin draft-dcbench-def-00 draft-bmwg-dcbench-methodology-01 1July 30 th, 2013IETF 87 – Berlin – BMWG

2. RFC 1242, 2432 (1991, 1998) RFC254 4 (1999) Network Interconnect Devices RFC2889 (2000) LAN Switching Devices Extending RFC2544 RFC391 8 (2004) Methodology for IP Multicast Benchmarking Latency, throughput and definitions ????? Virtualization, Multi- purpose/tenant DC, Low Latency, Big Data, Bursts, Buffering, Incast, Goodput… Evolution of the Data Center 2July 30 th, 2013IETF 87 – Berlin – BMWG-Avramov

3. RFC 1242, 2432 (1991, 1998) Benchmarking Terminology for Network Interconnect Devices, IP Multicast Benchmarking RFC2544 (1999) Network Interconnect Devices RFC2889 (2000) LAN Switching Devices Extending RFC2544 RFC3918 (2004) Methodology for IP Multicast Benchmarking DCBench Draft Redefine Terminology – Update Definitions Update methodology Implement with test vendors(Ixia) Community / Customer adoption Data Center Benchmarking is a very important need for us. For example, there is no standard and concrete methodology and implementation today to measure buffer utilization, efficiency and impact on the network equipment. Therefore there is no capability to understand and to master this critical network element. – Valérie Monte - Update DUT Benchmarking 3 July 30 th, 2013IETF 87 – Berlin – BMWG-Avramov

4. Redefine Definitions draft-dcbench-def-00 Redefine Definitions draft-dcbench-def-00 Latency Redefine how latency calculations are used Update usage of FIFO, FILO, LIFO and LILO Jitter Define the application Jitter RFC 3393 and packet size requirement and histogram for DC devices Physical Layer Calibration Cable test calibrations and documentation Line Rate Consequences of PPM: 99.98% Buffering Define Buffering and Buffer Efficiency, Burst, Intensity of Microburst Define Incast [many-one, many-many] Application Throughput Goodput definition and how to measure it Line Rate Testing Test all ports at 99.98% including latency, jitter histogram for min/max/avg and drops Buffering Testing Buffer highest efficiency Maximum port buffer size Maximum port pair buffer size Maximum DUT buffer size Microburst MicroBurst Testing Use all ports, at 100% intensity of microburst Head of Line Blocking Testing Measure two groups (8 ports) of DUT, up to all ports Reports provides percent of traffic loss during HOLB Incast Stateful and Stateless Traffic measure TCP goodput while measuring UDP latency Redefine Methodology draft-bmwg-dcbench-methodology-01 Redefine Methodology draft-bmwg-dcbench-methodology-01 Data Center Benchmarking Drafts Overview July 30 th, 2013 IETF 87 – Berlin – BMWG-Avramov4

5. 3.2.2 Measure Max Port Buffer Size This DUT has 4.29 MB of buffer per port at 1518 bytes with COS 0 Measured with Ixia July 30 th, 2013 IETF 87 – Berlin – BMWG-Avramov5

6. Conversation on draft-dcbench-def-00 Received great feedback on the list and have provided answers 1.2 Change measurement to measurement units 2.1 change Latency to Latency interval – FILO -> First bit, Last bit – FIFO -> First bit, First bit – LILO -> Last bit, Last bit – LIFO -> Last bit, First bit 2.3 Latency: FILO as MUST, FIFO as MAY 3. Jitter: what we mean by this is to provide new recommendation: packets of the same size MUST be used, and histogram of latency buckets MAY be provided 2.1 define cut-through additionally to bit forwarding [this is already in section 2.1] Provide buffering per priority [this is done as refered to cos /dscp value] July 30 th, 2013IETF 87 – Berlin – BMWG-Avramov6