Commit f397377b authored by Brewer, Wes's avatar Brewer, Wes
Browse files

Refined and document the hardware TDP and performance specs that were used for philly

parent cdff947d

config/philly/2-gpu.yaml

+5 −5
Original line number Diff line number Diff line
@@ -13,15 +13,15 @@ system: 
  cpus_per_node: 2

  cores_per_cpu: 20

  gpus_per_node: 2

  cpu_peak_flops: 1248000000000.0

  gpu_peak_flops: 7800000000000.0

  cpu_peak_flops: 1248000000000.0 # assume Xeon E5-2690v4 CPU 64-bit

  gpu_peak_flops: 9300000000000.0 # assume 12G P100 32-bit

  cpu_fp_ratio: 0.667

  gpu_fp_ratio: 0.667

power:

  power_gpu_idle: 75

  power_gpu_max: 300

  power_gpu_idle: 30

  power_gpu_max: 250

  power_cpu_idle: 90

  power_cpu_max: 280

  power_cpu_max: 270

  power_mem: 74.26

  power_nvme: 30

  power_nic: 20

config/philly/8-gpu.yaml

+5 −5
Original line number Diff line number Diff line
@@ -13,15 +13,15 @@ system: 
  cpus_per_node: 2

  cores_per_cpu: 20

  gpus_per_node: 8

  cpu_peak_flops: 1248000000000.0

  gpu_peak_flops: 7800000000000.0

  cpu_peak_flops: 1248000000000.0  # assume Xeon E5-2690v4 CPU 64-bit

  gpu_peak_flops: 12000000000000.0 # assume 24G P40 32-bit

  cpu_fp_ratio: 0.667

  gpu_fp_ratio: 0.667

power:

  power_gpu_idle: 75

  power_gpu_max: 300

  power_gpu_idle: 50

  power_gpu_max: 250

  power_cpu_idle: 90

  power_cpu_max: 280

  power_cpu_max: 270

  power_mem: 74.26

  power_nvme: 30

  power_nic: 20

raps/dataloaders/philly.py

+12 −1
Original line number Diff line number Diff line 
"""

Main reference to Philly traces:

This is the dataloader for the Philly traces which is documented in this paper:



    Jeon, Myeongjae, et al. "Analysis of Large-Scale Multi-Tenant GPU clusters for DNN training workloads." 

    2019 USENIX Annual Technical Conference (USENIX ATC 19). 2019.

    https://www.usenix.org/system/files/atc19-jeon.pdf



Note on hardware specs:



    Philly only provides GPU memory sizes (12G & 24G) without clarifying GPU models.

    Hu et al. (2024) https://arxiv.org/html/2403.07648v1



    For estimating system power and FLOPS performance, we assume that the 2-GPU

    nodes used Tesla P100 (12 GB) GPUs and the 8-GPU nodes used Tesla P40 (24 GB) 

    GPUs, consistent with hardware Microsoft deployed around 2017. Training is 

    assumed to have been performed in 32-bit (FP32), and the CPUs are assumed 

    to be 64-bit Intel Xeon E5-2690 v4.



The repository is available here:



    https://github.com/msr-fiddle/philly-traces