Neural personalized recommendation is the cornerstone of a wide collection of cloud services and products, constituting significant compute demand of cloud infrastructure. Thus, improving the execution efficiency of recommendation directly translates into infrastructure capacity saving. In this paper, we propose DeepRecSched, a recommendation inference scheduler that maximizes latency-bounded throughput by taking into account characteristics of inference query size and arrival patterns, model architectures, and underlying hardware systems. By carefully optimizing task versus data-level parallelism, DeepRecSched improves system throughput on server class CPUs by 2× across eight industry-representative models. Next, we deploy and evaluate this optimization in an at-scale production datacenter which reduces end-to-end tail latency across a wide variety of recommendation models by 30%. Finally, DeepRecSched demonstrates the role and impact of specialized AI hardware in optimizing system level performance (QPS) and power efficiency (QPS/watt) of recommendation inference. In order to enable the design space exploration of customized recommendation systems shown in this paper, we design and validate an end-to-end modeling infrastructure, DeepRecInfra. DeepRecInfra enables studies over a variety of recommendation use cases, taking into account at-scale effects, such as query arrival patterns and recommendation query sizes, observed from a production datacenter, as well as industry-representative models and tail latency targets.

Loop agreement is a family of distributed tasks that includes set agreement and simplex agreement, and was used to prove the undecidability of wait-free solvability of distributed tasks by read/write memory. Herlihy and Rajsbaum defined the algebraic signature of a loop agreement task, which consists of a group and a distinguished element. They used the algebraic signature to characterize the relative power of loop agreement tasks. In particular, they showed that one task implements another exactly when there is a homomorphism between their respective signatures sending one loop to the other. In this paper, we extend the previous result by defining the composition of multiple loop agreement tasks to create a new one with the same combined power. We generalize the original algebraic characterization for relative power to compositions of tasks. In this way, we can think of loop agreement tasks in terms of their basic building blocks. We also investigate a category-theoretic perspective of loop agreement by defining a category of loops, showing that the algebraic signature is a functor, and proving that our definition of task composition is the "correct" one, in a categorical sense.

MOTIVATION: Network alignment aims to find conserved regions between different networks. Existing methods aim to maximize total similarity over all aligned nodes (i.e. node conservation). Then, they evaluate alignment quality by measuring the amount of conserved edges, but only after the alignment is constructed. Thus, we recently introduced MAGNA (Maximizing Accuracy in Global Network Alignment) to directly maximize edge conservation while producing alignments and showed its superiority over the existing methods. Here, we extend the original MAGNA with several important algorithmic advances into a new MAGNA++ framework. RESULTS: MAGNA++ introduces several novelties: (i) it simultaneously maximizes any one of three different measures of edge conservation (including our recent superior [Formula: see text] measure) and any desired node conservation measure, which further improves alignment quality compared with maximizing only node conservation or only edge conservation; (ii) it speeds up the original MAGNA algorithm by parallelizing it to automatically use all available resources, as well as by reimplementing the edge conservation measures more efficiently; (iii) it provides a friendly graphical user interface for easy use by domain (e.g. biological) scientists; and (iv) at the same time, MAGNA++ offers source code for easy extensibility by computational scientists. AVAILABILITY AND IMPLEMENTATION: http://www.nd.edu/∼cone/MAGNA++/

MOTIVATION: Biological network alignment aims to identify similar regions between networks of different species. Existing methods compute node similarities to rapidly identify from possible alignments the high-scoring alignments with respect to the overall node similarity. But, the accuracy of the alignments is then evaluated with some other measure that is different than the node similarity used to construct the alignments. Typically, one measures the amount of conserved edges. Thus, the existing methods align similar nodes between networks hoping to conserve many edges (after the alignment is constructed!). RESULTS: Instead, we introduce MAGNA to directly 'optimize' edge conservation while the alignment is constructed, without decreasing the quality of node mapping. MAGNA uses a genetic algorithm and our novel function for 'crossover' of two 'parent' alignments into a superior 'child' alignment to simulate a 'population' of alignments that 'evolves' over time; the 'fittest' alignments survive and proceed to the next 'generation', until the alignment accuracy cannot be optimized further. While we optimize our new and superior measure of the amount of conserved edges, MAGNA can optimize any alignment accuracy measure, including a combined measure of both node and edge conservation. In systematic evaluations against state-of-the-art methods (IsoRank, MI-GRAAL and GHOST), on both synthetic networks and real-world biological data, MAGNA outperforms all of the existing methods, in terms of both node and edge conservation as well as both topological and biological alignment accuracy. AVAILABILITY: Software: http://nd.edu/∼cone/MAGNA CONTACT: : [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.