Energy e ciency is becoming more important in database
design, i. e., the work delivered by a database server should
be accomplished by minimal energy consumption. So far, a
substantial number of research papers examined and
optimized the energy consumption of database servers or single
components. In this way, our rst e orts were exclusively
focused on the use of ash memory or SSDs in a DBMS context
to identify their performance potential for typical DB
operations. In particular, we developed tailor-made algorithms to
support caching for ash-based databases [
A key observation made by Tsirogiannis et al. [
Because the energy use of single-server systems is far from being energy proportional, we came up with the hypothesis that better energy e ciency may be achieved by a cluster of nodes whose size is dynamically adjusted to the current workload demand. For this reason, we shifted our research focus from in exible single-server DBMSs to distributed clusters running on lightweight nodes. Although distributed systems impose some performance degradation compared to a single, brawny server, they o er higher energy saving potential in turn.
Current hardware is not energy proportional, because a
single server consumes, even when idle, a substantial
fraction of its peak power [
We have developed a research prototype of a distributed DBMS called WattDB on a scale-out architecture, consisting of n wimpy computing nodes, interconnected by an 1GBit/s Ethernet switch. The cluster currently consists of 10 identical nodes, composed of an Intel Atom D510 CPU, 2 GB DRAM and an SSD. The con guration is considered Amdahl-balanced, i. e., balanced between I/O and network throughput on one hand and processing power on the other.
Compared to In niBand, the bandwidth of the interconnecting network is limited but su cient to supply the lightweight nodes with data. More expensive, yet faster connections would have required more powerful processors and more sophisticated I/O subsystems. Such a design would have pushed the cost beyond limits, especially because we would not have been able to use commodity hardware. Furthermore, by choosing lightweight components, the overall energy footprint is low and the smallest con guration, i. e., the one with the fewest number of nodes, exhibits low power consumption. Moreover, experiments running on a small cluster can easily be repeated on a cluster with more powerful nodes.
A dedicated node is the master node, handling incoming queries and coordinating the cluster. Some of the nodes have each four hard disks attached and act as storage nodes, providing persistent data storage to the cluster. The remaining nodes (without hard disks drives) are called processing nodes. Due to the lack of directly accessible storage, they can only operate on data provided by other nodes (see Figure 1).
All nodes can evaluate (partial) query plans and execute DB operators, e. g., sorting, aggregation, etc., but only the storage nodes can access the DB storage structures, i. e., tables and indexes. Each storage node maintains a DB bu er
S S D
S S D
S S D S S D
S S D S S D S S D
to keep recently referenced pages in main memory, whereas a processing node does not cache intermediate results. As a consequence, each query needs to always fetch the quali ed records from the corresponding storage nodes.
Hence, our cluster design results in a shared-nothing architecture where the nodes only di erentiate to those which have or have not direct access to DB data on external storage. Each of the nodes is additionally equipped with a 128GB Solid-State Disk (Samsung 830 SSD). The SSDs do not store the DB data, they provide swap space to support external sorting and to provide persistent storage for conguration les. We have chosen SSDs, because their access latency is much lower compared to traditional hard disks; hence, they are better suited for temp storage.
In WattDB, a dedicated component, running on the master node, controls the energy consumption, called EnergyController. This component monitors the performance of all nodes in the cluster. Depending on the current query workload and node utilization, the EnergyController activates and suspends nodes to guarantee a su ciently high node utilization depending on the workload demand. Suspended nodes do only consume a fraction of the idle power, but can be brought back online in a matter of a few seconds. It also modi es query plans to dynamically distribute the current workload on all running nodes thereby achieving balanced utilization of the active processing nodes.
As data-intensive workloads, we submit speci c TPC-H queries against a distributed shared-nothing DBMS, where time and energy use are captured by speci c monitoring and measurement devices. We con gure various static clusters of varying sizes and show their in uence on energy e ciency and performance. Further, using an EnergyController and a load-aware scheduler, we verify the hypothesis that energy proportionality for database management tasks can be well approximated by dynamic clusters of wimpy computing nodes. 1.