Skip to content

xiandong79/Interference-Aware-Cluster-Management

Repository files navigation

Interference Aware Cluster Management

Background

  1. Paragon: QoS-Aware Scheduling for Heterogeneous Datacenters - ASPLOS ’13
  2. Quasar: Resource-Efficient and QoS-Aware Cluster Management - ASPLOS ’14
  1. SPEC CPU2006 单线程负载
  2. memcached 内存型数据库
  3. parsec 多线程负载
  4. websearch Latency critical 任务
  5. perf/lib-perf 任务性能检测
  1. bash_basic.sh - SPEC CPU 2006 任务间相互干扰
  2. memcached+spec2006.sh - memcached+spec2006 任务间相互干扰
  3. ibench.sh autorun_ibench.sh - SPEC CPU 2006 任务在 ibench 七种不同压力干扰下的运行状况

Data Analysis

  1. ALS_SGD_MF.py
# Train rmse: 0.632234683903
# Test rmse: 0.958863923627
  1. gridsearch_ALS_SGD_MF.py - 可遍历地求出最优超参数

  2. 基于服务器 IPS参数 的推荐系统-20170823版本.html

载入原始数据¶
In [2]:
# Load data from disk

names = ['workload_id', 'pressure_id', 'rating']
df = pd.read_csv('/Users/dong/Desktop/体系-数据分析/IPS-rating-final.csv',delimiter=",", names=names)

print(df.shape)

num_workloads = df.workload_id.unique().shape[0]
num_pressures = df.pressure_id.unique().shape[0]

print(num_workloads, "kinds of workloads")
print(num_pressures, "kinds of pressures")
(86, 3)
12 kinds of workloads
8 kinds of pressures

Results and Conclusion

在未来的使用中,每次任务提交时,只需在IPS-rating-final.csv文件中,继续补充 此种workload_id 的在 pressure_id 测试值(2-3次),即可得出此种 workload在每一种压力下的 “百分制评分”。 Greedily选择最高评分即可。

Prediction Result...
[[ 53  81  52  74 100  96  99 101]
 [ 49  79  49  72  98  94  97 100]
 [ 50  79  48  71  99  95  98 100]
 [ 51  78  47  70  97  93  97  99]
 [ 50  78  49  71  97  94  97  99]
 [ 50  78  50  71  96  93  95  98]
 [ 51  78  47  70  97  93  97  99]
 [ 86  81  58  76  88  87  92  95]
 [ 49  80  51  73 100  96  99 101]
 [ 50  79  50  72  98  94  97 100]
 [ 69  80  70  77  87  85  87  88]
 [ 79  80  78  79  82  81  82  83]]