Datasets for MONet: Heterogeneous Memory over Optical Network for Large-Scale Data Centre Resource Disaggregation

Fig. 4 MONet: Switch-Plane Characterization - Architecture Power and Latency

Switch Plane Characterization: Power and network latency comparison between Non-Parallel (fat tree) and MONet architectures

Fig. 5: MONet: Remote memory access Round Trip Latency

DDR4/HMC local/remote (8m) memory read/write latency: 8-bonded transceivers each at 10,12.5, 15 Gb/s

Fig 6: MONet: DDR4/HMC Remote memory read/write latency overhead

Remote memory read/write latency: Impact of optical distance b/w CPU and remote memory on round-trip latency. Values are measured experimentally for 8, 18 and 36m; only 100m is based on Eq1.

Fig 7. MONet DDR4: Achieved Bandwidth, Memory/Link Utilization

MONet DDR4 Access: Achieved bandwidth, link and memory bandwidth utilization for locally and remotely (8m) attached DDR4: transceiver lanes at rates (10, 12.5 and 15 Gb/s)

Fig 8. MONet HMC: Achieved Bandwidth, Memory/Link Utilization

Achieved bandwidth, link and memory bandwidth utilization for locally and remotely (8 m) attached HMC, compared with achieved maximum memory bandwidth for different lane rates (10, 12.5, 15 Gb/s)

Fig. 9 MONet: Power Consumption Distribution

Power consumption distribution between CPU and memory over 8-metre round-trip optical data path. Round-trip net energy efficiency (with and without MONet’s resources) and memory-to-link ratio over number of transceivers link and lane rate.

Fig. 10: MONet HMC Access: Physical Layer Performance

Physical layer performance of a single bi-directional channel CPU and HMC: Received optical power (dBm) vs log10(BER).

Fig. 11: MONet HMC Access: BER vs Bandwidth

Impact of Bit Error Rate (BER) on memory bandwidth performance per one HMC half width link (8 transceivers).

Fig. 12: MONet STREAM Benchmark DDR4

STREAM benchmark performance for DDR4 at 8-metres round-trip distance using 8 and single channel

Fig. 13: MONet HMC STREAMS benchmark

Application level performance using the STREAM benchmark for accessing serial memory (local and remote at 8 metres round-trip) at 10, 12.5 and 15 Gb/s lane rate.

Fig. 14 MONet: DDR4 and HMC: STREAM and baseline

Sustained STREAM and baseline bandwidth (8-links) over round-trip over round-trip optical distance: 8, 18, 26,36 metres.

Fig. 15-16: MONet DDR4: Memcached Throughput

Achieved Throughput in Workload (A, B, C and F) when DDR4 is locally/remotely attached. For DDR4: parallel accessed (MM), stream data-width size in bytes (8 to 64). Sustained Throughput in Workload (A, B, C and F)when DDR4 is remotely attached at round-trip optical distance 8, 16, 26 and 36-metres.

Fig. 15-16: MONet HMC: Memcached Throughput

Achieved Throughput in Workload (A, B, C and F) when HMC is locally/remotely attached. For HMC: full-width (FW) (16-lane) and half-width (HW) (8-lane)at 10, 12.5 and 15 Gb/s bit-rates. Sustained Throughput in Workload (A, B, C and F) when HMC is remotely attached at round-trip optical distance 8, 16, 26 and 36-metres.

Fig. 17-18 MONet: DDR4 Memcached Latency

Achieved Average Latency in Workload (A, B, C and F) when DDR4 is locally and remotely attached. For local attachment in DDR4: parallel accessed (MM), stream data-width size in bytes (8 to 64). Sustained Average Latency in Workload (A, B, C and F) when DDR4 is remotely attached at round-trip optical distance 8, 16, 26 and 36-metres

Fig. 17-18: MONet HMC Memcached Latency

Achieved Average Latency in Workload (A, B, C and F) when HMC is locally and remotely attached. For local attachment in HMC: Full-width (FW) (16-lane) and half-width (HW) (8-lane) at 10, 12.5 and 15 Gb/s bit-rates. Sustained Average Latency in Workload (A, B, C and F) when HMC is remotely attached at round-trip optical distance 8, 16, 26 and 36-metres

Fig. 19: MONet DDR4 Memcached IPC

Impact of optical distance on IPC in workload (A,B, C and F) for whole CPU. For DDR4: using 8 and one transceivers links each at 10, 12.5 and 15 Gb/s.

Fig. 19: MONet HMC Memcached IPC

Impact of optical distance on IPC in workload (A,B, C and F) for whole CPU. For HMC: Half-width (8-lane) at 10, 12.5 and 15 Gb/s bit-rates