Improving Consolidation of Virtual Machines with Risk-aware Bandwidth Oversubscription in Compute Clouds Amir Epstein Joint work with David Breitgand 1 © 2009 IBM Corporation Motivation Network Bandwidth is a critical Data Center resource Network Bandwidth may become a bottleneck for consolidation Accurate and efficient network bandwidth demand estimation is difficult Common practice: fully provision for peak loads Consequences: resource waste 2 © 2009 IBM Corporation Full Provisioning VS. Multiplexing The aggregate demand of VMs may be much smaller than the sum of the maximum demand of each VM: ∑i maxt di(t) >> maxt ∑i di(t) 70 60 Capacity 50 40 VM1 VM1-Max 30 20 10 0 1 10 19 28 37 46 55 64 73 82 91 Max(VM1)+Max(VM2)=110 100 Time 60 50 Capacity 40 VM2 VM2-Max 30 20 10 0 1 10 19 28 37 46 55 64 73 82 91 100 Time 3 © 2009 IBM Corporation Full Provisioning VS. Multiplexing 80 70 60 Capacity 50 VM1 VM2 40 VM1+VM2 Max: VM1+VM2 30 20 10 0 1 11 21 31 41 51 61 71 81 91 Time Max(VM1+VM2)=71 < Max(VM1)+Max(VM2)=110 4 © 2009 IBM Corporation Statistical Multiplexing Consider each VM dynamic bandwidth demands as a random variable Consider the aggregate bandwidth demand which is a sum of the random variables representing VMs Bandwidth demands As the number of VMs increases: – The ratio between standard deviation of the aggregate bandwidth demand and the mean decreases 5 © 2009 IBM Corporation Overcommit Cloud provider aims at improving cost-efficiency Overcommit resources using statistical multiplexing Our focus is bandwidth 6 © 2009 IBM Corporation Stochastic Bin Packing Problem (SBP) S={X1,…, Xn} – Set of items Xi – random variable representing the size (bandwidth demand) of item i p – overflow probability Goal: Partition the set S into the smallest number of subsets (bins) S1,…,Sk such that Pr[ X i: X i S j i 1] p for 1 j k p represents a probabilistic SLA / policy 7 © 2009 IBM Corporation SBP with Normal Distribution We assume that each item i independently follows normal distribution N(μi ,σi2) . When σi,=0, for all i, then Xi= μi and the problem reduces to the classical bin packing problem The focus of this work is SBP with normal variables 8 © 2009 IBM Corporation Related Work – Bin Packing The problem is NP-hard Bin packing is hard to approximate to a factor better than 3/2 unless P=NP. First Fit Decreasing (FFD) has asymptotic approximation ratio of 11/9 and (absolute) approximation ratio of 3/2. MFFD algorithm has asymptotic approximation ratio of 71/60. AFPTAS exists. Online bin packing – First Fit (FF) has competitive ratio of 17/10. – Best upper and lower bounds are 1.58899 and 154014, respectively. 9 © 2009 IBM Corporation Related Work – Stochastic Bin Packing log p 1 O 1 log log p -approximation for SBP with Bernoulli variables [Kleinberg et. al 1997] SBP with Poisson, Exponential and Bernoulli variables [Goel and Indik 1999] – PTAS exists for Poisson and exponential distributions. – Quasi-PTAS exists for Bernoulli variables. – These results relax bin capacity and overflow probability constraints by a factor 1+ε. (1 2)(1 ) - competitive algorithm for SBP with normal variables [Wang et. al 2011] 10 © 2009 IBM Corporation Our Results 2-approximation algorithm for SBP with normal variables (2+ε)-competitive algorithm for online SBP with normal variables Observe the existence of a dual PTAS for SBP with normal variables. 11 © 2009 IBM Corporation Definitions Definition: The effective load of bin j is l j i: X i S j i i: X i S j i2 where 1 (1 p) and the quantile function 1 is the inverse function of the CDF Ф of N(0,1). Observation: A packing is feasible for a given overflow probability p iff for every bin j, lj i: X i S j i i: X i S j i2 1 The load of bin j is normally distributed with mean i and i: X i S j variance i2 i: X i S j 12 © 2009 IBM Corporation Simple solution approach Reduce the problem to the classical bin packing problem with item sizes i i , thus P( X i i i ) 1 p A feasible solution to the classical bin packing problem is a feasible solution SBP, since i: X i S j i i: X i S j 2 i i: X i S j ( i i ) 1 The optimum for the classical bin packing instance with the new sizes may be significantly larger than the optimum for SBP. 13 © 2009 IBM Corporation Effective Size l j i 2 i iS j i iS j iS j ( i ) 2 iS j 2 i Thus, the effective size of item i on bin j can be viewed as i ( i ) 2 2 i iS j 14 © 2009 IBM Corporation Approximation Algorithm Algorithm 1: First Fit VMR decreasing Order the items in non-increasing order of VMR Place the next item in the first bin into which it can be feasibly packed If no such bin exists, open a new bin to pack this item Variance to Mean Ratio (VMR) is 15 d i / i 2 i © 2009 IBM Corporation Approximation Algorithm Theorem 1: Algorithm 1 is a 2-approximation algorithm for SBP with normal variables. 16 © 2009 IBM Corporation Integer Program for SBP n xij i i 1 m x j 1 ij 1 x ij {0,1} 17 n 2 x ij i 1 1 j m, i 1 1 i n, 1 i n, 1 j m © 2009 IBM Corporation Mathematical Program Relaxation n xij i i 1 m x j 1 ij 1 x ij 0 18 n 2 x ij i 1 1 j m, i 1 1 i n, 1 i n, 1 j m © 2009 IBM Corporation Fractional Algorithm (Algorithm 2) Order the items in non-increasing order of VMR Place the next item in the bin with remaining capacity. If the item causes an overflow to the bin, assign maximum fraction of this item to the bin. Then, open a new bin to pack the remaining part of this item. Variance to Mean Ratio (VMR) is 19 di i2 / i © 2009 IBM Corporation Analysis Lemma: There exists a feasible solution to the MP with the following property. For any pair of items k,l and a pair of bins i<j, if xkj>0 and xli>0, then dl ≥ dk. Observation: Fractional algorithm produces a feasible fractional solution to the MP. This implies that collocating items with high VMR (bursy) minimizes the total effective size of the items Variance to Mean Ratio (VMR) is 20 d i / i 2 i © 2009 IBM Corporation Proof Outline Consider a feasible solution to the MP with lexicographically maximal standard deviation (STD) vector of the bins i2 S=(S1,…,Sm), where S j i: X S Assume by contradiction that the items are not packed into the bins according to non-increasing order of VMR Thus, there exists at least one pair of items that are not placed in this order (i.e., item with smaller VMR is packed to a bin with smaller index than the other item). We show that we can exchange fractions of these items between the bins, such that – the new solution is feasible – The STD vector of the bins in the solution is lexicographically greater than the one in the original solution 21Contradiction i j © 2009 IBM Corporation Online Algorithm VMR di i / i Let 1 8 1 C ln log1 4 2 Class 0: di 2 2 k 1 2 k (1 ) d (1 ) Class 1≤k≤C: i Class C+1: 1/ 2 2 (1 )C di 22 © 2009 IBM Corporation Online Algorithm Algorithm 3: Classify next item according to the VMR classes Place the next item in the first bin of its class into which it can be feasibly packed If no such bin exists, open a new bin to pack this item Theorem 2: Algorithm 3 is a (2+O(ε))-approximation algorithm for SBP with normal variables. 23 © 2009 IBM Corporation Simulation Study Compare our proposed algorithms to previous reported ones Data set – Real trace from production data center used to compute mean and standard deviation of bandwidth consumption of 6000 VMs over a few hours period. – Synthetic traces with statistical properties similar to those of the real traces 24 © 2009 IBM Corporation Algorithms Algorithms 1-3 First Fit (FF) with deterministic item sizes μi+βσi First Fit Decreasing (FFD) with deterministic item sizes μi+βσi Group Packing (GP) [Wang et. al 2011] For the online algorithms (Algorithm 3 and Group Packing), we set ε=0.1. 25 © 2009 IBM Corporation Real Instance (Online) 26 (Approx.) (L.B) © 2009 IBM Corporation Real Instance (Online) 27 (Approx.) (L.B) © 2009 IBM Corporation Real Instance (Online) 28 (Approx.) (L.B) © 2009 IBM Corporation Online Algorithms Large synthetic instances 9% 8% 29 © 2009 IBM Corporation Summary We studied SBP under the assumption that virtual machines bandwidth demand obeys normal distribution We showed a 2-approximation algorithm We showed (2+ε)-competitive algorithm We observed the existence of a dual PTAS for SBP We studied the performance and applicability of our algorithms using synthetic and real data The performance evaluation showed that our proposed algorithms considerably reduce the number of bins compared to the best known algorithms for the problem 30 © 2009 IBM Corporation