A Fair Comparison of VM Placement Heuristics and a More Effective Solution

Scalable correlation-aware virtual machine consolidation using two-phase clustering

“…Considerable efforts have been made in the VM placement problem in recent years, aiming at minimizing the amount of servers required or energy consumption [15], [18]- [20]. In order to guarantee the promised SLA, the interference among co-located VMs has also been studied, usually through analyzing low-level hardware usage metrics and workload characteristics [6]- [8], as well as consolidation approaches built on top of the interference models [3], [9], [10].…”

“…• ScalCCon multi : considers both, CPU and RAM, in the VM placement by integrating RBP, which is our prior VM placement algorithm using a fixed buffer to handle utilization peaks [18]. The idea is to demonstrate 1) the application of ScalCCon in multiple dimensional scenarios; 2) through the comparison with RBP that the size of the fixed reserved buffer can be reduced, so that resources can be utilized more efficiently by leveraging correlation analysis.…”

“…The first VM that can fit in the currently considered PM is selected from the first Peak Cluster (i.e., the Peak Cluster with the largest peak) (line 6). VMs in the studied Peak Cluster are sorted by CPU in ScalCCon cpu , or by cosine similarity as used in our previously proposed VM placement algorithm RBP [18] (i.e., by how similar the VM's shape is with that of the current Algorithm 1 ScalCCon VM Placement Algorithm 1: pmList ← sortP M s()//by CPU in descending order 2: peakClusterList ← sortP eakClusters() //by maximum peak of clusters in descending order 3: for all pm in pmList do 4: //Pick the first VM to be placed 5: sort VMs in the first Peak Cluster pCluster //by CPU or cosine similarity in descending order 6: place the first vm that can fit in pm or the vm having the most similar shape with pm 7: if pCluster has only one sub-cluster then 8: selectedP Cluster ← pCluster + 1 //move to next peak Cluster 9: peakBuf f er ← vm s peak 10: else 11: selectedP Cluster ← pCluster 12: peakBuf f er ← maximum peak of pCluster 13: end if 14: //Fill pm with VMs that are least correlated 15: while pm is not full and vmList.size() > 0 do 16: select the sub-cluster least correlated with pm 17: sort VMs in the selected sub-cluster 18: select the vm having the biggest CPU or the most similar shape with pm 19: if selectedP Cluster has only one sub-cluster then 20: selectedP Cluster ← selectedP Cluster + 1 //move to next Peak Cluster 21: end if 22: end while 23: if vmList == N U LL then 24: break 25: end if 26: end for PM's residual capacity) in ScalCCon multi (line 5). The size of a VM in terms of CPU is estimated as the 90th percentile of the VM's CPU usage over the time series.…”

SOC: Satisfaction-Oriented Virtual Machine Consolidation in Enterprise Data Centers

“…SOC is flexible to integrate any VM placement algorithm that takes a set of PMs and a set of VMs as inputs. In this paper, we employ an effective algorithm RBP [14], which minimizes the number of PMs required, and the resource wastage on each allocated PM by placing VMs that are complementary to PMs' residual capacity.…”

“…Different from the conventional matchmaking, the input of SOC 's matchmaking process is the manager's preferences instead of demand or job descriptions. The second task is to apply Resource Balancing Placement (RBP), a VM placement algorithm proposed in our previous work [14], to allocate VMs to PMs.…”