Deploying large language models (LLMs) on edge devices presents significant challenges due to the substantial computational overhead and memory requirements.Activation sparsification can mitigate these resource challenges by reducing the number of activated neurons during inference.Existing methods typically employ thresholding-based sparsification based on the statistics of activation tensors.However, they do not model the impact of activation sparsification on performance, resulting in suboptimal performance degradation.To address the limitations, this paper reformulates the activation sparsification problem to explicitly capture the relationship between activation sparsity and model performance.Then, this paper proposes CHESS , a general activation sparsification approach via CHannel-wise thrEsholding and Selective Sparsification.First, channel-wise thresholding assigns a unique threshold to each activation channel in the feed-forward network (FFN) layers.Then, selective sparsification involves applying thresholding-based activation sparsification to specific layers within the attention modules.Finally, we detail the implementation of sparse kernels to accelerate LLM inference.Experimental results demonstrate that the proposed CHESS achieves lower performance degradation over eight downstream tasks while activating fewer parameters than existing methods, thus speeding up the LLM inference by up to 1.27x.

The relatively high cost of record deserialization is increasingly becoming the bottleneck of column-based storage systems in tree-structured applications [58]. Due to record transformation in the storage layer, unnecessary processing costs derived from fields and rows irrelevant to queries may be very heavy in nested schemas, significantly wasting the computational resources in large-scale analytical workloads. This leads to the question of how to reduce both the deserialization and IO costs of queries with highly selective filters following arbitrary paths in a nested schema. We present CORES (Column-Oriented Regeneration Embedding Scheme) to push highly selective filters down into column-based storage engines, where each filter consists of several filtering conditions on a field. By applying highly selective filters in the storage layer, we demonstrate that both the deserialization and IO costs could be significantly reduced. We show how to introduce fine-grained composition on filtering results. We generalize this technique by two pair-wise operations, rollup and drilldown, such that a series of conjunctive filters can effectively deliver their payloads in nested schema. The proposed methods are implemented on an open-source platform. For practical purposes, we highlight how to build a column storage engine and how to drive a query efficiently based on a cost model. We apply this design to the nested relational model especially when hierarchical entities are frequently required by ad hoc queries. The experiments, including a real workload and the modified TPCH benchmark, demonstrate that CORES improves the performance by 0.7×--26.9× compared to state-of-the-art platforms in scan-intensive workloads.

As the reasoning aspects and the knowledge based processing capabilities of RDF (Resource Description Framework) have been widely adopted in W3C Recommendation, the ontology layer and query languages of the Semantic Web stack achieve a certain level of maturity. There exists an increasing need for high performance, read-only semantic analysis for the massive RDF data. In this demo we will present a Map-Reduce based SPARQL processing engine, called SHOE (SPARQL on Hadoop with Optimization Encoding), to handle billions of RDF triples. SHOE consists of three major components (1) the RDF data loader, (2) the partition generator and (3) the query processor. While this demonstration mainly focuses on enhancing SPARQL processing in the Hadoop platform, the underlying encoding and partitioning optimization strategies can be utilized by the common Map-Reduce frameworks in the share-nothing environment.

In the era of "big data", the emergence and increasing adoptions of the related enabling technologies make it possible for Map-Reduce to accommodate DSS (Decision Support Systems) load, which is commonly targeted for high-performance Data Warehouse analyses in the context of RDBMS. However, the non-predetermined mapping of the Map-Reduce tasks to the physical machines makes it difficult to utilize the pre-partitioned and indexing techniques of DBMS to improve the data locality. In this paper, towards multi-way join evaluating OLAP (Online Analysis Processing) workloads, we introduce table partitioning by reference to Map-Reduce. For avoiding the dispersion of the initial tuples that belong to the same segment keys, we present a detailed description of the data organization model that partitions the dominated tables by cascade reference constraints. In order to push multiple joins on these clustered partitions down to the map task, we design a one-pass multi-way join algorithm along with its optimization implementations for the major Map-Reduce stages. We conduct an empirically study with TPCH benchmark on different scales of clusters, and experimentally verify the high efficiency of the proposed optimization model.

This paper presents a run-time scheduling policy to map tasks to resources in grid computing systems based on Multi-Agent System (MAS). This policy schedules tasks in run-time and avoids the waste of resources. A simulation result is presented to prove the availability of this policy.