{"id":19,"date":"2021-02-03T20:36:48","date_gmt":"2021-02-04T01:36:48","guid":{"rendered":"https:\/\/sites.bu.edu\/casp\/?page_id=19"},"modified":"2026-10-04T10:48:14","modified_gmt":"2026-10-04T14:48:14","slug":"research","status":"publish","type":"page","link":"https:\/\/sites.bu.edu\/casp\/research\/","title":{"rendered":"Research"},"content":{"rendered":"

Our research spans on multiple areas of data-centric systems. Recently, we have been working on developing practical solutions for private collaborative analytics with Multi-Party Computation<\/a>, designing adaptive systems for data stream processing<\/a>, and\u00a0building systems for graph Machine Learning<\/a> that are scalable, reliable, and work on streaming data. We also work on performance modeling for cloud-native applications<\/a>.<\/p>\n

Systems for Secure Computation<\/h3>\n
\n
\n
\n

Research in the cryptography community has shown that any function can be securely computed by a group of participants in a distributed fashion such that each party learns the function’s output and nothing more. This fascinating idea started as a theoretical curiosity in the 1980s but since then it has evolved into a useful tool to realize use cases like the gender wage gap study<\/a> in the Boston area. A series of recent theoretical advances in the field have further supported the belief that provably secure computation can perhaps become as practical and ubiquitous as public-key cryptography. Some of these advances include efficient multi-party computation protocols, fully homomorphic encryption schemes, zero-knowledge proofs, and oblivious RAM. Supporting end-to-end secure computation with practical performance requires rethinking the entire systems stack<\/strong>: from the programming abstractions all the way down to the hardware. Below are two recent results of our research in this area.<\/p>\n

A unified engine for secure analytics<\/strong><\/p>\n

CryptDough<\/strong><\/a> (SOSP’26) is our most recent system in this line of work. Existing MPC systems typically support a single type of workload or a single threat model. CryptDough brings relational analytics, time series computations, and ML inference into one engine, so users can combine them into end-to-end secure data pipelines. It supports multiple MPC protocols across the full range of threat models within the same runtime. CryptDough also introduces virtual vectors<\/em>, which let users write simple single-threaded code while the engine handles communication, parallelization, and memory management. In our experiments, CryptDough matches the functionality of state-of-the-art MPC systems and is often more than 2\u00d7 faster on the workloads they support. See the paper<\/a> and try out the code<\/a>.<\/p>\n

\"\"<\/a><\/p>\n

Secure relational analytics at scale<\/strong><\/p>\n<\/div>\n

ORQ<\/strong><\/a> (SOSP’25), the\u00a0Oblivious Relational Query engine<\/em>, significantly improves the performance and expressivity of SECRECY for secure collaborative analytics in the outsourced setting. We introduce a new join operator and implement more efficient sorting protocols which allow us to evaluate unique-key joins in O<\/em>(n <\/em>log n<\/em>) time, rather than\u00a0O<\/em>(n<\/em>\u00b2). Query rewriting under relational algebra then allows the vast majority of practical queries, even those with arbitrary joins, to be evaluated by this operator. ORQ runs the entire TPC-H benchmark, for the first time under MPC, and executes all queries from prior work significantly faster and on orders-of-magnitude larger inputs.<\/p>\n

ORQ received the Distinguished Artifact Award<\/strong> at SOSP’25, and an extended version was invited for publication in ACM TOCS<\/a>. ORQ has also been used in practice: it powered the first Massachusetts Workforce Data Report<\/a>. See the paper<\/a> and code<\/a>.<\/p>\n

Earlier work<\/strong><\/p>\n

ORQ and CryptDough build on two earlier systems. SECRECY<\/a> (NSDI’23) was the first to offer privacy-preserving collaborative analytics as a cloud service, with a cost-based optimization framework for relational MPC that does not rely on trusted infrastructure. TVA<\/a> (Security’23) extended this approach to time series analytics, supporting oblivious time windows, keyed aggregations, and filters while keeping all data attributes private, including timestamps.<\/p>\n

\n

Funding and next steps<\/strong><\/p>\n<\/div>\n<\/div>\n

Our research on system for secure computation is supported by several new awards. John Liagouris<\/a> received an NSF CAREER award<\/a> for the project “A Unified Analytics Stack for Secure Computation<\/strong>” and an Amazon Research Award<\/a> for “Pushing secure MPC beyond niche applications<\/strong>“.\u00a0 More recently, PI Mayank Vari<\/a>a and co-PIs John Liagoguris<\/a> and Vasiliki Kalavri<\/a> received a 3-year, $1.2M award from the NSF CICI program<\/a> to build OPAL<\/strong>, the Open Privacy-preserving Analytics Lab<\/strong>: an open-source platform for running privacy-preserving scientific workflows across open and commercial clouds, using MPC and trusted execution environments (TEEs). OPAL will serve both as a testbed for organizations to evaluate these tools and as an orchestration engine for real-world, multi-cloud deployments.<\/div>\n
\n
\n

We have also recently received an NSF CAREER award<\/a> that will allow us to conduct fundamental research towards developing a novel streaming dataflow software stack, tailored to the heterogeneity of emerging workloads and computing environments.<\/p>\n

\n

<\/h3>\n<\/div>\n

Adaptive Data Stream Processing Systems<\/h3>\n

The success of open-source streaming dataflow systems, like Apache Flink and Spark Streaming, has generated a vibrant ecosystem of applications, including online content recommendation, continuous traffic monitoring, complex event processing, anomaly detection, and data enrichment for online model predictions. However, the increasing diversity of workloads combined with the emergence of compute-enabled edge devices, hardware accelerators, and new cloud computing paradigms, is pushing these decade-old frameworks to their limits. Our goal is to develop next-generation stream processing systems that treat adaptability, resource efficiency, and platform independence as first-class concerns, and that are easily accessible to non-expert users.<\/p>\n<\/div>\n

Reconfiguration without disruption<\/strong><\/p>\n

Koala<\/a><\/strong> (NSDI’27) is a new reconfiguration protocol<\/em> for long-running stateful dataflow systems that retains the operational performance of embedded-state architectures while matching the disruption-free adaptability of external-state designs. Unlike prior protocols which relocate all state regardless of whether it will soon be accessed, Koala migrates state in a fine-grained, lazy manner. To achieve that, it employs a lightweight ownership transfer mechanism that safely decouples event routing from state location. Koala will appear at NSDI’27 next year but you can already start using it! See the code<\/a>.<\/p>\n

\n

\"\"<\/a><\/p>\n<\/div>\n<\/div>\n

\n
\n

 <\/p>\n<\/div>\n

We have also recently received an NSF CAREER award<\/a> that will allow us to conduct fundamental research towards developing a novel streaming dataflow software stack, tailored to the heterogeneity of emerging workloads and computing environments.<\/p>\n

Earlier work<\/strong><\/p>\n

CAPSys<\/a> (EuroSys’25) is a contention-aware task placement approach for stream processing [code]<\/a>. We also recently collaborated with the Apache Flink community on Disaggregated State Management in Apache Flink 2.0<\/a> (VLDB’25). Before that, we quantified the cost of auto-scaling for non-expert users (BiDEDE’22<\/a>), built a benchmark harness for streaming state stores (EuroSys’22<\/a>), and argued for workload-aware state management (HotStorage’20<\/a>).<\/p>\n

This research is supported by Vasiliki Kalavri’s NSF CAREER award<\/a> on a novel streaming dataflow software stack tailored to the heterogeneity of emerging workloads and computing environments.<\/p>\n

\n

Systems for Graph Machine Learning<\/h3>\n

Graph Neural Networks (GNNs) are a powerful model for machine learning on graph-structured data, with applications in recommendation, fraud detection, search, molecular biology, and code summarization. Real-world graphs are massive, dynamic, and often incomplete, and existing GNN systems require distributed computation, large amounts of memory, and multiple GPUs. Our work addresses three questions: how to give users formal guarantees on the answers that ML-powered graph queries return, how to run GNN inference on streaming data, and how to train GNNs at scale on a single machine.<\/p>\n

\"\"<\/a><\/p>\n

GNN inference on streaming data<\/strong><\/p>\n

Many graphs change continuously and applications need predictions as new data arrives. We received an i-Scale award<\/a> for the project “Streaming GNN Inference for Real-time Anomaly Detection on Microservice Graphs<\/strong>“, which explores using GNNs to improve security monitoring in cloud-native microservice environments. This builds on our GRADES-NDA’25 paper<\/a>, which outlines the challenges and opportunities of combining GNN inference with dataflow stream processing [code<\/a>].<\/p>\n

\n

Reliable ML-powered graph queries<\/strong><\/p>\n

Neural graph databases fill in missing links in incomplete knowledge graphs using ML predictors. This is powerful, but prediction errors can compound across multi-hop queries, with no bound on how many true answers are missed. ConRAD<\/a> (VLDB’26), a collaboration with KTH, lets users set a risk budget and uses Conformal Risk Control to automatically derive per-operator thresholds that meet the recall target with formal guarantees, while maximizing precision. It also introduces a conformal gate<\/em> operator that skips neural inference when the graph itself provides enough evidence, which cuts unnecessary model calls in well-connected regions of the graph.<\/p>\n

\n

Scaling GNN training on modern storage<\/strong><\/p>\n

With support from an NSF award<\/a>, we are designing systems for static and dynamic GNNs that target a single commodity machine with SSD storage, as a cost-effective alternative to distributed GNN architectures. RingSampler<\/a> (HotStorage’25) uses Linux io_uring for efficient GNN sampling on large graphs stored on SSDs [code<\/a>]. Our earlier DaMoN’24 paper<\/a> (Best Short Paper Award<\/strong>) uses near-storage compute and a high-throughput FPGA kernel to pipeline sampling across training epochs [code<\/a>].<\/p>\n<\/div>\n<\/div>\n","protected":false},"excerpt":{"rendered":"

Our research spans on multiple areas of data-centric systems. Recently, we have been working on developing practical solutions for private collaborative analytics with Multi-Party Computation, designing adaptive systems for data stream processing, and\u00a0building systems for graph Machine Learning that are scalable, reliable, and work on streaming data. We also work on performance modeling for cloud-native […]<\/p>\n","protected":false},"author":18854,"featured_media":0,"parent":0,"menu_order":0,"comment_status":"closed","ping_status":"closed","template":"","meta":[],"_links":{"self":[{"href":"https:\/\/sites.bu.edu\/casp\/wp-json\/wp\/v2\/pages\/19"}],"collection":[{"href":"https:\/\/sites.bu.edu\/casp\/wp-json\/wp\/v2\/pages"}],"about":[{"href":"https:\/\/sites.bu.edu\/casp\/wp-json\/wp\/v2\/types\/page"}],"author":[{"embeddable":true,"href":"https:\/\/sites.bu.edu\/casp\/wp-json\/wp\/v2\/users\/18854"}],"replies":[{"embeddable":true,"href":"https:\/\/sites.bu.edu\/casp\/wp-json\/wp\/v2\/comments?post=19"}],"version-history":[{"count":50,"href":"https:\/\/sites.bu.edu\/casp\/wp-json\/wp\/v2\/pages\/19\/revisions"}],"predecessor-version":[{"id":1166,"href":"https:\/\/sites.bu.edu\/casp\/wp-json\/wp\/v2\/pages\/19\/revisions\/1166"}],"wp:attachment":[{"href":"https:\/\/sites.bu.edu\/casp\/wp-json\/wp\/v2\/media?parent=19"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}