Awesome

Abacus

This repository contains the source code for a research paper that was submitted for publication at the International Conference for High Performance Computing, Networking, Storage, and Analysis (SC21).

What is Abacus

Abacus is a runtime system that runs multiple DNN queries simultaneously with stable and predictable latency. Abacus enables deterministic operator overlap to enforce the latency predictability. Abacus is comprised of an overlap-aware latency predictor, a headroom-based query controller, and segmental model executors. The latency predictor is able to precisely predict the latencies of queries when the operator overlap is determined. The query controller determines the appropriate operator overlap to guarantee the QoS of all the DNN services on a GPU. The model executors run the operators as needed to support the deterministic operator overlap. Our evaluation using seven popular DNNs on an Nvidia A100 GPU shows that Abacus significantly reduces the QoS violation and improves the throughput compared with state-of-the-art solutions.

Environment Preparation

• Hardware&software requirements

  1. Hardware Requirements

     1. CPU: Intel(R) Xeon(R) Silver 4210R CPU @ 2.40GHz
     2. Memroy: 252G
     3. NVIDIA Ampere 100

  2. Software Requirements

     1. Ubuntu 20.04.1 (Kernel 5.8.0)
     2. GPU Driver: 460.39
     3. CUDA 11.2
     4. CUDNN 8.1
     5. Anaconda3-2020.7
     6. Pytorch 1.8.0

• Preparing Python environment

  1. Install Anaconda3 as the Python runtime

  $ cd ~ && mkdir -p .local && cd .local
  $ wget -O Anaconda3-2020.07-Linux-x86_64.sh https://repo.anaconda.com/archive/Anaconda3-2020.07-Linux-x86_64.sh
  $ chmod +x Anaconda3-2020.07-Linux-x86_64.sh
  $ ./Anaconda3-2020.07-Linux-x86_64.sh -b -p ../.local/anaconda3
  

  2. Activate conda and create python environment with essential dependencies

  $ eval "$($HOME/.local/anaconda3/bin/conda shell.zsh hook)"
  $ # cd into Abacus repository
  $ conda create --name Abacus
  $ pip3 install torch==1.7.1+cu110 torchvision==0.8.2+cu110 -f https://download.pytorch.org/whl/torch_stable.html
  $ pip3 install -r requirements.txt
  

  3. Nvidia GPU related affairs: switch MIG and MPS

     • MIG: The new Multi-Instance GPU (MIG) feature allows GPUs based on the NVIDIA Ampere architecture (such as NVIDIA A100) to be securely partitioned into up to seven separate GPU Instances for CUDA applications, providing multiple users with separate GPU resources for optimal GPU utilization.
     • MPS: The MPS runtime architecture is designed to transparently enable co-operative multi-process CUDA applications

     We provide a shell scripts for configuring the MIG and MPS related features on the NVIDIA GPUs. The detailed usage is as follows:

  $ chmod +x nvidia_utils.sh
  $ # switch mig, 0: disable, 1: enable
  $ ./nvidia_utils.sh --mig 0/1
  $ # create GPU instances
  $ ./nvidia_utils.sh --cgi 1/2/4
  $ # delete GPU instances
  $ ./nvidia_utils.sh --dgi
  $ # switch mps with mig disabled, options: start/stop
  $ ./nvidia_utils.sh --mps start/stop
  $ # switch mps with mig enabled, options: start/stop
  $ ./nvidia_utils.sh --mps start/stop --mig 1/2/4
  

Getting Started

The following sections step through the things required to run Abacus

Profiling

Abacus needs to profile the essential data for training a precise overlap-aware latency predictor.

We first profiling the data without MPS enabled and MIG disabled.

• Profiling for pair-wise co-location on a dedicated A100.

  $ python main.py --task profile --model_num 2
  

• Profiling for triplet-wise co-location on a dedicated A100.

  $ python main.py --task profile --model_num 3
  

• Profiling for quadruplet-wise co-location on a dedicated A100.

  $ python main.py --task profile --model_num 4
  

We then profiling the data without MPS and MIG both enabled.

• Profiling for pair-wise co-location on a MIG 2g.10gb of A100.

  $ python main.py --task profile --model_num 2 --mig 2
  

• Profiling for quadruplet-wise co-location on a MIG 4g.10gb of A100.

  $ python main.py --task profile --model_num 4 --mig 1
  

Training Predictor

After obataining all the profiling data, we train the latency predictor for each cases.

Training MLP model

• Training the predictor for pair-wise co-location on a dedicated A100 for each combination.

  $ python main.py --task train --model_num 2 --mode one by one --modeling mlp
  

• Training the predictor for pair-wise co-location on a dedicated A100 for all combinations.

  $ python main.py --task train --model_num 2 --mode all --modeling mlp
  

• Training the predictor for triplet-wise co-location on a dedicated A100 for all combinations.

  $ python main.py --task train --model_num 3 --mode all --modeling mlp
  

• Training the predictor for quadruplet-wise co-location on a dedicated A100 for all combinations.

  $ python main.py --task train --model_num 4 --mode all --modeling mlp
  

• Training the predictor for pair-wise co-location on a MIG 2g.10gb of A100 for all combinations.

  $ python main.py --task train --model_num 2 --mode all --modeling mlp --mig 2
  

• Training the predictor for quadruplet-wise co-location on a MIG 4g.20gb of A100 for all combinations.

  $ python main.py --task train --model_num 4 --mode all --modeling mlp --mig 1
  

Training LR/SVM model

• Training the predictor for pair-wise co-location on a dedicated A100 for all combinations.

  $ python main.py --task train --model_num 2 --mode all --modeling lr/svm
  

• Training the predictor for pair-wise co-location on a dedicated A100 for each combination.

  $ python main.py --task train --model_num 2 --mode one by one --modeling lr/svm
  

5.5 Determining Modeling Techniques

We can get the prediction error from the output in terminal after training the predictor with MLP/LR/SVM models. To get the cross-validation results, we only need to re-train the model because the random seed for generating the dataset is automatically changed.

We organize the data of errors as shown in data/modeling/2in7results.csv.We also provide a script experiments/5.5_prediction/plot.py for plot the results.

The following figure depicts the prediction errors for all modeling methods.

Online Serving

After profiling and training, we can serve multiple DNN services with Abacus

• Testing Abacus for pair-wise co-location on a dedicated A100, take the setup of model id: 0, 1; qos_target: 50ms; total tested queries: 1000; search ways: 2 as an example.

  $ python main.py --task serve --model_num 2 --comb 0 1 --policy Abacus --load 50 --qos 50 --queries 1000 --thld 5 --ways 2 --abandon
  

• Testing Abacus for triplet-wise co-location on a dedicated A100, take the setup of model id: 0, 1, 3; qos_target: 50ms; total tested queries: 1000; search ways: 2 as an example.

  $ python main.py --task serve --model_num 2 --comb 0 1 2 --policy Abacus --load 50 --qos 50 --queries 1000 --thld 5 --ways 2 --abandon
  

• Testing Abacus for quadruplet-wise co-location on a dedicated A100, take the setup of model id: 0, 1, 2, 3; qos_target: 50ms; total tested queries: 1000; search ways: 2 as an example.

  $ python main.py --task serve --model_num 2 --comb 0 1 2 3 --policy Abacus --load 50 --qos 50 --queries 1000 --thld 5 --ways 2 --abandon
  

• Testing Abacus for pair-wise co-location on a MIG 2g.10gb of A100, take the setup of model id: 0, 1; qos_target: 50ms; total tested queries: 1000; search ways: 2 as an example.

  $ python main.py --task serve --model_num 2 --comb 0 1 --policy Abacus --load 50 --qos 50 --queries 1000 --thld 5 --ways 2 --abandon --mig 2
  

• Testing Abacus for pair-wise co-location on a MIG 4g.20gb of A100, take the setup of model id: 0, 1, 2, 3; qos_target: 50ms; total tested queries: 1000; search ways: 2 as an example.

  $ python main.py --task serve --model_num 2 --comb 0 1 --policy Abacus --load 50 --qos 50 --queries 1000 --thld 5 --ways 2 --abandon --mig 1
  

Evaluation

All evaluations are conducted in the root directory of Abacus. The following table presents the detailed information of each experiments, including the corresponding figures in paper and the shell scripts for running the experiment.

7.2 Ensuring QoS

• We first evaluate the QoS guarantee in terms of tail latency and QoS violation ratio.

  # Normal
  $ ./experiments/7.2_qos/2in7.sh
  # Small DNNs
  $ ./experiments/7.2_qos/smalldnn.sh
  

  The following figure depicts the 99%-ile latency for each pair-wise co-location.

  <img src="figure/2in7_qos.png" width="1000">
  End-to-end 99%-ile latency of each pair-wise co-location with FCFS, SJF, EDF, and Abacus.

  The following two figures depict the QoS violation ratio for each pair-wise co-location and the 99%-ile latency for small DNNs.

  <img src="figure/qos_violation_ratio.png" width="400">	<img src="figure/small_dnn.png" width="400">
  QoS violation ratio of each pair-wise co-location with<br /> FCFS, SJF, EDF, and Abacus.	99%-ile latencies with small co-located DNNs

7.3 Improving Peak Throughput

• We then evaluate the peak throughput under the load that exceeds the hardware limit.

  $ ./experiments/7.3_throughput/2in7.sh
  

  The following figure depicts the peak throughput for each pair-wise co-location.

  <img src="figure/2in7_throughput.png" width="1000"/>
  The peak throughput of each co-location pair with FCFS, SJF, EDF, and Abacus while guaranteeing the QoS.

7.4 Beyongd Pair-wise Co-location

• We also evaluate Abacus beyond pair-wise co-location.

  # QoS
  $ ./experiments/7.4_beyond_pair/qos/3in4.sh
  $ ./experiments/7.4_beyond_pair/qos/4in4.sh
  # Throughput
  $ ./experiments/7.4_beyond_pair/throughput/3in4.sh
  $ ./experiments/7.4_beyond_pair/throughput/3in4.sh
  

  The following two figures presents the performance of Abacus for triplet-wise and quadruplet-wise co-location in terms of QoS guarantee and peak throughput.

  <img src="figure/3in4_qos.png" width="580">	<img src="figure/3in4_throughput.png" width="580">
  The 99%-ile latency in triplets- and quadruplets- wise deployments with FCFS, SJF, EDF, and Abacus.	Peak throughputs in triplets- and quadruplets- wise deployments with FCFS, SJF, EDF, and Abacus.

7.5 Integrating with MIGs

• We also evaluate Abacus with MIG enabled.

  # QoS
  $ ./experiments/7.5_mig/qos/1in4.sh
  $ ./experiments/7.5_mig/qos/2in4.sh
  $ ./experiments/7.5_mig/qos/4in4.sh
  # Throughput
  $ ./experiments/7.5_mig/throughput/1in4.sh
  $ ./experiments/7.5_mig/throughput/2in4.sh
  $ ./experiments/7.5_mig/throughput/3in4.sh
  

  The following two figures presents the performance of Abacus for MIG 1g.10gb and MIG 2g.20gb in terms of QoS guarantee and peak throughput.

7.6 Applying in a DNN Serving Cluster

 # Client
 $ ./experiments/7.6_cluster/client.sh
 # Server
 $ ./experiments/7.6_cluster/server.sh

The following figure presents

7.7 Effectiveness of Multi-way Search

• We also evaluate the effectiveness of multi-way search The latency is given while training the MLP model

$ python main.py --task train --model_num 2 --mode all --modeling mlp