Skip to main content

Advances, Systems and Applications

Table 1 Comparison of data centre RM approaches

From: Integrated resource management pipeline for dynamic resource-effective cloud data center

Authors Architecture Target When to migrate Performance Metric Conclusion
Nathuji and Centralized minimize power - Periodic reallocation - Power (W) The overall energy
Schwan [12]   consumption without    consumption could be reduced significantly, up
   performance penalties    to 34%, without appreciable losses in performance.
Verma et al. [13] Centralized minimize power - Periodic reallocation - Power (watts) pMapper was considered
   consumption,   - Energy (kilojoules) an efficient solution to
   considering the   - Migration cost minimize power
   VM migration   - Overall cost consumption (less than
   cost   - Power savings (%) 0.2% penalty).
Zhu et al. [15] De-centralized improve - Periodic reallocation - Response time The integration of node
  - Pod controllers workload - Threshold-based heuristics (seconds) and pod controllers
  - Node controllers management to ensure efficient use of data center resources   - Number of migrations improved performance by 32% and 23% over fixed allocation and over non-integrated controllers, and reduced migrations for high priority workloads.
Gmach et al. [16] De-centralized minimize power - Periodic reallocation - Migration overhead The integration between
  - workload placement consumption, - Threshold-based heuristics - CPU quality reactive migration
  controller taking into   violations controller and periodic
  - migration controller account the Qos and the number of VM migrations   - Power consumption workload placement controller presented the best approach for power and SLA, but needs more migrations.
VMware Distributed Centralized minimize power - Fixed threshold heuristics   Fixed threshold heuristics
Power Management   consumption    are unsuitable for real
(DPM) [17]      systems with dynamic and unknown workloads.
Li et al. [18] Centralized minimize power - Dynamic threshold heuristics - Energy The double threshold with
   consumption and QoS   - Number of migrations multi-resource utilization with the MPSO algorithm
   guarantee   - Number of active physical servers reduces energy consumption and
     - load balance degree improves the QoS.
Beloglazov et Centralized minimize power - Fixed threshold heuristics - Energy (kWh) It is not a suitable decision
al. [24]   consumption,   - SLAVs (%) for keeping the utilization
   taking into account QoS   - Number of migrations threshold constant as the workload is in continuous change.
Beloglazov and De-centralized minimize power - Decision based on statistical - Energy (kWh) The proposed LR
Buyya [23]   consumption, analysis of historical data - ESV algorithm remarkably
   taking into   - SLAVs outperformed other
   account QoS   - PDM (%) dynamic VM consolidation
     - Number of migrations algorithms.
Guenter et al. [21] De-centralized minimize power - Decision based on statistical - Power saving Predicting demand was
   consumption, analysis of historical data - normalized daily used to switch on servers
   considering the   energy savings before require and avoid
   trade-off between cost, performance,   (MWh) switching on unnecessary servers.
   and reliability    
Bobroff et al. [22] Centralized minimize power - Decision based on statistical - Time-averaged The proposed algorithm
   consumption analysis of historical data number of servers used decreased the
     used number of servers needed
     - Capacity of overflow to support a certain SLA by 50% compared to static consolidation.
This work De-centralized minimize power - Fixed threshold heuristics - AITF (%) Our combination of DES,
   consumption, respect - Decision based on statistical - AOTF (%) MMTMC 2, and MF
   overall and end-user’s analysis of historical data - Number of migrations algorithms improved
   SLA, and eliminate   - Energy saving (%) performance in power
   unnecessary migrations    saving, QoS, and network traffic. The number of migrations reduced by 49.44% compared to default algorithms.