Task placement for Minimizing Total Power in Heterogeneous Datacenters

This project deals with the problem of task placement in a heterogeneous datacenter to minimize total power (computational and cooling power) consumption which is directly related to the operational cost of the datacenter. Cross Interference based task placement algorithm XInt-h is already proposed by IMPACT Lab in [2] where a GA based approach was taken to find the near optimal task placement for the objective of minimizing maximum Inlet Temperature of the servers in the data center. In this project the GA based algorithm for the XInt-h approach has been formulated for the objective of minimizing the Total Power. The problem of minimizing the total operation cost through heterogeneous task placement is formulated as a discrete minimax problem of non-linear equations with linear constraints. CFD simulations have been used to get recirculation characteristics in the datacenter and power measurements from real data center equipments have been used to provide a comparative analysis of the minimax and GA formulation of the XInt-h approach with the MinHR approach. Results show that the formulation outperforms MinHR, and in some cases it can half the total operation cost of the data center with respect to a task placement by MinHR. Also simulations are performed with the different objectives of minimizing total power as well as minimizing cooling power and the results show that only at low utilizations the decision on the objective function matters.

Current trends in data center growth show a rapid increase in both computing and storage capacity. For a large data center the annual cost can run into millions of dollars and the cooling cost accounts for almost 35% of the operation cost [1] . Computer Room Air Conditioning (CRAC) units operate at a much lower temperature than the equipment’s redline temperature (typically 25–35 °C), to compensate for the effects of heat recirculation. Heat recirculations should be avoided as they give rise to hot spots. However according to the operational characteristics of cooling systems, operating at a higher temperature is more energy-efficient, because the co-efficient of performance is better. To address this problem Impact Lab presents a formulation of this problem in terms of minimax optimization and provides simulation for a comparative analysis of this algorithm with the MinHR task placement algorithm.

The datacenter model was simulated in Flovent 6.1 a CFD simulator. The CFD simulation of the datacenter was necessary in order to get the recirculation coefficients of the datacenter. The recirculation coefficient matrix of the datacenter is obtained through a number of profiling steps that can be performed via actual measurements in the datacenter or via CFD simulation of intelligently devised scenarios. The model of datacenter on which the simulations are performed is shown in figure 1. Due to the limited simulation capability of the CFD software, we simulated a small scale data center with physical dimensions 9.6m × 8.4m × 3.6 m, which has two rows of industry standard 42U racks arranged in a typical cold aisle and hot aisle layout. The cold air is supplied by one computer room air conditioner, with the flow rate 8m3/s. The cold air rises from raised floor plenum through vent tiles, and exhausted hot air returns to the air conditioner through ceiling vent tiles. There are 10 racks and each rack is equipped with 5 chassis. Each 7U (12.25-inch) chassis can hold 10 blade servers. We assume that 20 nodes are node Type 1 (CPU-intensive) and the remaining 30 nodes are Type 2 (IO-intensive). Two type of applications running inside data center would have four different power consumption profiles on 2 types of nodes.. In addition, within utilized nodes, two applications take 30% and 70% of server resource, respectively. We prepared different data center layouts: For example, it is possible the first n1 nodes are of Type 1 and the remaining n2 node are of Type 2, or the upper part of nodes are type 1 and lower part of nodes are type 2, etc. A test case scenario showing the recirculation of heat is given in figure 2 and the temperature map of the datacenter is given in figure 3.

The problem formulation is based on the Mathematical model of the data center proposed by Impact Lab [2].

• a data center of n nodes, each node i having:
  • m_i processors (cores), and
  • an idle chassis power consumption of w_i;
• p tasks already running on the server with:
  • C₀ their n × p placement table;
• q tasks, each demanding c(k) processors (cores);
• a composite task-server power profile matrix B º {b_ik}_{n ×(p+q)};
• the heat distribution matrix D;

Simulations were performed in Flovent 6.1 and Matlab environment. The data center was modelled using Flovent 6.1 a CFD simulator. From CFD simulations we obtained the heat recirculation coefficients and then used these coefficients to simulate the effect of task placement in the datacenter according to different algorithms.
The first set of simulations compared the minimax formulation of the cross interference based task placement approach XInt-h with its GA formulation and with the MinHR approach with objective of minimizing the total power. From figure 4 it is clear that the XInt-h approach outperform the MinHR approach for the objective of producing low power task placements. Figure 5 which gives the Supply temperature required to keep all the inlet temperatures below red line for different utilization also shows that the XInt-h approach is better than the MinHR approach in terms of Cooling Power and efficiency. The SHI vs Utilization plot supports the conjecture.

To improve the energy efficiency and the reliability of data center operation, thermal-aware workload placement or placement has been studied to improve temperature distribution within data centers. We analytically expressed the task placement’s effect on the data center energy needs and formulated a minimax optimization problem for task placement in a heterogeneous environment. This minimax formulation of the XInt-h algorithm is proved to provide better task placements in heterogenous datacenter than MinHR.

• [1] Jeffrey Rambo, Yogendra Joshi ''Modeling of data center airflow and heat transfer: State of the art and future trends", Published Online: January 19th 2007 Springer Science + Business Media, LLC 2007

• [2] Q. Tang, and S. K. S. Gupta ''Thermal Aware Task Scheduling for Datacenters through Minimizing Heat Recirculation", Cluster 2007, Austin TX. Sept. 2007.