Physical Testbeds

All FD.io CSIT performance test results included in this report are executed on the physical testbeds hosted by LF FD.io project, unless otherwise noted.

Two physical server topology types are used:

  • 2-Node Topology: Consists of one server acting as a System Under Test (SUT) and one server acting as a Traffic Generator (TG), with both servers connected into a ring topology. Used for executing tests that require frame encapsulations supported by TG.

  • 3-Node Topology: Consists of two servers acting as a Systems Under Test (SUTs) and one server acting as a Traffic Generator (TG), with all servers connected into a ring topology. Used for executing tests that require frame encapsulations not supported by TG e.g. certain overlay tunnel encapsulations and IPsec. Number of native Ethernet, IPv4 and IPv6 encapsulation tests are also executed on these testbeds, for comparison with 2-Node Topology.

Current FD.io production testbeds are built with SUT servers based on the following processor architectures:

  • Intel Xeon: Skylake Platinum 8180, Haswell-SP E5-2699v3, Cascade Lake Platinum 8280, Cascade Lake 6252N.

  • Intel Atom: Denverton C3858.

  • Arm: TaiShan 2280, hip07-d05.

  • AMD EPYC: Zen2 7532.

Server SUT performance depends on server and processor type, hence results for testbeds based on different servers must be reported separately, and compared if appropriate.

Complete technical specifications of compute servers used in CSIT physical testbeds are maintained in FD.io CSIT repository: https://git.fd.io/csit/tree/docs/lab/testbed_specifications.md.

Following is the description of existing production testbeds.

2-Node AMD EPYC Zen2 (2n-zn2)

One 2n-zn2 testbed in in operation in FD.io labs. It is built based on two SuperMicro SuperMicro AS-1114S-WTRT servers, with SUT and TG servers equipped with one AMD EPYC Zen2 7532 processor each (256 MB Cache, 2.40 GHz, 32 cores). 2n-zn2 physical topology is shown below.

testbed-2n-zn2

SUT server is populated with the following NIC models:

  1. NIC-1: x710-DA4 4p10GE Intel.

  2. NIC-2: xxv710-DA2 2p25GE Intel.

  3. NIC-3: cx556a-edat ConnectX5 2p100GE Mellanox.

TG server runs TRex application and is populated with the following NIC models:

  1. NIC-1: x710-DA4 4p10GE Intel.

  2. NIC-2: xxv710-DA2 2p25GE Intel.

  3. NIC-3: cx556a-edat ConnectX5 2p100GE Mellanox.

All AMD EPYC Zen2 7532 servers run with AMD SMT enabled, doubling the number of logical cores exposed to Linux.

2-Node Xeon Cascade Lake (2n-clx)

Three 2n-clx testbeds are in operation in FD.io labs. Each 2n-clx testbed is built with two SuperMicro SYS-7049GP-TRT servers, SUTs are equipped with two Intel Xeon Gold 6252N processors (35.75 MB Cache, 2.30 GHz, 24 cores). TGs are equiped with Intel Xeon Cascade Lake Platinum 8280 processors (38.5 MB Cache, 2.70 GHz, 28 cores). 2n-clx physical topology is shown below.

testbed-2n-clx

SUT servers are populated with the following NIC models:

  1. NIC-1: x710-DA4 4p10GE Intel.

  2. NIC-2: xxv710-DA2 2p25GE Intel.

  3. NIC-3: cx556a-edat ConnectX5 2p100GE Mellanox.

  4. NIC-4: empty, future expansion.

  5. NIC-5: empty, future expansion.

  6. NIC-6: empty, future expansion.

TG servers run T-Rex application and are populated with the following NIC models:

  1. NIC-1: x710-DA4 4p10GE Intel.

  2. NIC-2: xxv710-DA2 2p25GE Intel.

  3. NIC-3: cx556a-edat ConnectX5 2p100GE Mellanox.

  4. NIC-4: empty, future expansion.

  5. NIC-5: empty, future expansion.

  6. NIC-6: x710-DA4 4p10GE Intel. (For self-tests.)

All Intel Xeon Cascade Lake servers run with Intel Hyper-Threading enabled, doubling the number of logical cores exposed to Linux.

2-Node Xeon Skylake (2n-skx)

Four 2n-skx testbeds are in operation in FD.io labs. Each 2n-skx testbed is built with two SuperMicro SYS-7049GP-TRT servers, each in turn equipped with two Intel Xeon Skylake Platinum 8180 processors (38.5 MB Cache, 2.50 GHz, 28 cores). 2n-skx physical topology is shown below.

testbed-2n-skx

SUT servers are populated with the following NIC models:

  1. NIC-1: x710-DA4 4p10GE Intel.

  2. NIC-2: xxv710-DA2 2p25GE Intel.

  3. NIC-3: empty, future expansion.

  4. NIC-4: empty, future expansion.

  5. NIC-5: empty, future expansion.

  6. NIC-6: empty, future expansion.

TG servers run T-Rex application and are populated with the following NIC models:

  1. NIC-1: x710-DA4 4p10GE Intel.

  2. NIC-2: xxv710-DA2 2p25GE Intel.

  3. NIC-3: empty, future expansion.

  4. NIC-4: empty, future expansion.

  5. NIC-5: empty, future expansion.

  6. NIC-6: x710-DA4 4p10GE Intel. (For self-tests.)

All Intel Xeon Skylake servers run with Intel Hyper-Threading enabled, doubling the number of logical cores exposed to Linux, with 56 logical cores and 28 physical cores per processor socket.

3-Node Xeon Skylake (3n-skx)

Two 3n-skx testbeds are in operation in FD.io labs. Each 3n-skx testbed is built with three SuperMicro SYS-7049GP-TRT servers, each in turn equipped with two Intel Xeon Skylake Platinum 8180 processors (38.5 MB Cache, 2.50 GHz, 28 cores). 3n-skx physical topology is shown below.

testbed-3n-skx

SUT1 and SUT2 servers are populated with the following NIC models:

  1. NIC-1: x710-DA4 4p10GE Intel.

  2. NIC-2: xxv710-DA2 2p25GE Intel.

  3. NIC-3: empty, future expansion.

  4. NIC-4: empty, future expansion.

  5. NIC-5: empty, future expansion.

  6. NIC-6: empty, future expansion.

TG servers run T-Rex application and are populated with the following NIC models:

  1. NIC-1: x710-DA4 4p10GE Intel.

  2. NIC-2: xxv710-DA2 2p25GE Intel.

  3. NIC-3: empty, future expansion.

  4. NIC-4: empty, future expansion.

  5. NIC-5: empty, future expansion.

  6. NIC-6: x710-DA4 4p10GE Intel. (For self-tests.)

All Intel Xeon Skylake servers run with Intel Hyper-Threading enabled, doubling the number of logical cores exposed to Linux, with 56 logical cores and 28 physical cores per processor socket.

3-Node Xeon Haswell (3n-hsw)

Three 3n-hsw testbeds are in operation in FD.io labs. Each 3n-hsw testbed is built with three Cisco UCS-c240m3 servers, each in turn equipped with two Intel Xeon Haswell-SP E5-2699v3 processors (45 MB Cache, 2.3 GHz, 18 cores). 3n-hsw physical topology is shown below.

testbed-3n-hsw

SUT1 and SUT2 servers are populated with the following NIC models:

  1. NIC-1: VIC 1385 2p40GE Cisco.

  2. NIC-2: NIC x520 2p10GE Intel.

  3. NIC-3: empty.

  4. NIC-4: NIC xl710-QDA2 2p40GE Intel.

  5. NIC-5: NIC x710-DA2 2p10GE Intel.

  6. NIC-6: QAT 8950 50G (Walnut Hill) Intel.

TG servers run T-Rex application and are populated with the following NIC models:

  1. NIC-1: NIC xl710-QDA2 2p40GE Intel.

  2. NIC-2: NIC x710-DA2 2p10GE Intel.

  3. NIC-3: empty.

  4. NIC-4: NIC xl710-QDA2 2p40GE Intel.

  5. NIC-5: NIC x710-DA2 2p10GE Intel.

  6. NIC-6: NIC x710-DA2 2p10GE Intel. (For self-tests.)

All Intel Xeon Haswell servers run with Intel Hyper-Threading disabled, making the number of logical cores exposed to Linux match the number of 18 physical cores per processor socket.

2-Node Atom Denverton (2n-dnv)

2n-dnv testbed is built with: i) one Intel S2600WFT server acting as TG and equipped with two Intel Xeon Skylake Platinum 8180 processors (38.5 MB Cache, 2.50 GHz, 28 cores), and ii) one SuperMicro SYS-E300-9A server acting as SUT and equipped with one Intel Atom C3858 processor (12 MB Cache, 2.00 GHz, 12 cores). 2n-dnv physical topology is shown below.

testbed-2n-dnv

SUT server have four internal 10G NIC port:

  1. P-1: x553 copper port.

  2. P-2: x553 copper port.

  3. P-3: x553 fiber port.

  4. P-4: x553 fiber port.

TG server run T-Rex software traffic generator and are populated with the following NIC models:

  1. NIC-1: x550-T2 2p10GE Intel.

  2. NIC-2: x550-T2 2p10GE Intel.

  3. NIC-3: x520-DA2 2p10GE Intel.

  4. NIC-4: x520-DA2 2p10GE Intel.

The 2n-dnv testbed is in operation in Intel SH labs.

3-Node Atom Denverton (3n-dnv)

One 3n-dnv testbed is built with: i) one SuperMicro SYS-7049GP-TRT server acting as TG and equipped with two Intel Xeon Skylake Platinum 8180 processors (38.5 MB Cache, 2.50 GHz, 28 cores), and ii) one SuperMicro SYS-E300-9A server acting as SUT and equipped with one Intel Atom C3858 processor (12 MB Cache, 2.00 GHz, 12 cores). 3n-dnv physical topology is shown below.

testbed-3n-dnv

SUT1 and SUT2 servers are populated with the following NIC models:

  1. NIC-1: x553 2p10GE fiber Intel.

  2. NIC-2: x553 2p10GE copper Intel.

TG servers run T-Rex application and are populated with the following NIC models:

  1. NIC-1: x710-DA4 4p10GE Intel.

3-Node ARM TaiShan (3n-tsh)

One 3n-tsh testbed is built with: i) one SuperMicro SYS-7049GP-TRT server acting as TG and equipped with two Intel Xeon Skylake Platinum 8180 processors (38.5 MB Cache, 2.50 GHz, 28 cores), and ii) one Huawei TaiShan 2280 server acting as SUT and equipped with one hip07-d05 processor (64* ARM Cortex-A72). 3n-tsh physical topology is shown below.

testbed-3n-tsh

SUT1 and SUT2 servers are populated with the following NIC models:

  1. NIC-1: connectx4 2p25GE Mellanox.

  2. NIC-2: x520 2p10GE Intel.

TG server runs T-Rex application and is populated with the following NIC models:

  1. NIC-1: x710-DA4 4p10GE Intel.

  2. NIC-2: xxv710-DA2 2p25GE Intel.

  3. NIC-3: xl710-QDA2 2p40GE Intel.

2-Node ARM ThunderX2 (2n-tx2)

One 2n-tx2 testbed is built with: i) one SuperMicro SYS-7049GP-TRT server acting as TG and equipped with two Intel Xeon Skylake Platinum 8180 processors (38.5 MB Cache, 2.50 GHz, 28 cores), and ii) one Marvell ThnderX2 9975 (28* ThunderX2) server acting as SUT and equipped with two ThunderX2 ARMv8 CN9975 processors. 2n-tx2 physical topology is shown below.

testbed-2n-tx2

SUT server is populated with the following NIC models:

  1. NIC-1: xl710-QDA2 2p40GE Intel (not connected).

  2. NIC-2: xl710-QDA2 2p40GE Intel.

TG server run T-Rex application and is populated with the following NIC models:

  1. NIC-1: xl710-QDA2 2p40GE Intel.