arXiv subjects

Tadeusz Kobus

Publications and source records attributed to Tadeusz Kobus.

5 recordsLinked to original sources

arXiv2025

UPMEM Unleashed: Software Secrets for Speed

Developing kernels for Processing-In-Memory (PIM) platforms poses unique challenges in data management and parallel programming on limited processing units. Although software development kits (SDKs) for PIM, such as the UPMEM SDK, provide essential tools, these emerging platforms still leave significant room for performance optimization. In this paper, we reveal surprising inefficiencies in UPMEM software stack and play with non-standard programming techniques. By making simple modifications to the assembly generated by the UPMEM compiler, we achieve speedups of 1.6-2x in integer addition and 1.4-5.9x in integer multiplication, depending on the data type. We also demonstrate that bit-serial processing of low precision data is a viable option for UPMEM: in INT4 bit-serial dot-product calculation, UPMEM can achieve over 2.7x speedup over the baseline. Minor API extensions for PIM allocation that account for the non-uniform memory access (NUMA) architecture of the server further improve the consistency and throughput of host-PIM data transfers by up to 2.9x. Finally, we show that, when the matrix is preloaded into PIM, our optimized kernels outperform a dual-socket CPU server by over 3x for INT8 generalized matrix-vector multiplication (GEMV) and by 10x for INT4 GEMV. Our optimized INT8 GEMV kernel outperforms the baseline 3.5x.

cs.AR↗

arXiv2021

Jiffy: A Lock-free Skip List with Batch Updates and Snapshots

In this paper we introduce Jiffy, the first lock-free, linearizable ordered key-value index that offers both (1) batch updates, which are put and remove operations that are executed atomically, and (2) consistent snapshots used by, e.g., range scan operations. Jiffy is built as a multiversioned lock-free skip list and relies on CPU's Time Stamp Counter register to generate version numbers at minimal cost. For faster skip list traversals and better utilization of the CPU caches, key-value entries are grouped into immutable objects called revisions. Moreover, by changing the size of revisions and thus modifying the synchronization granularity, our index can adapt to varying contentions levels (smaller revisions are more suited for write-heavy workloads whereas large revisions benefit read-dominated workloads, especially when they feature many range scan operations). Structure modifications to the index, which result in changing the size of revisions, happen through (lock-free) skip list node split and merge operations that are carefully coordinated with the update operations. Despite rich semantics, Jiffy offers highly scalable performance, which is comparable or exceeds the performance of the state-of-the-art lock-free ordered indices that feature linearizable range scan operations. Compared to its (lock-based) rivals that also support batch updates, Jiffy can execute large batch updates up to 7.4x more efficiently.

cs.DS↗

arXiv2021

On Mixing Eventual and Strong Consistency: Acute Cloud Types

In this article we study the properties of distributed systems that mix eventual and strong consistency. We formalize such systems through acute cloud types (ACTs), abstractions similar to conflict-free replicated data types (CRDTs), which by default work in a highly available, eventually consistent fashion, but which also feature strongly consistent operations for tasks which require global agreement. Unlike other mixed-consistency solutions, ACTs can rely on efficient quorum-based protocols, such as Paxos. Hence, ACTs gracefully tolerate machine and network failures also for the strongly consistent operations. We formally study ACTs and demonstrate phenomena which are neither present in purely eventually consistent nor strongly consistent systems. In particular, we identify temporary operation reordering, which implies interim disagreement between replicas on the relative order in which the client requests were executed. When not handled carefully, this phenomenon may lead to undesired anomalies, including circular causality. We prove an impossibility result which states that temporary operation reordering is unavoidable in mixed-consistency systems with sufficiently complex semantics. Our result is startling, because it shows that apparent strengthening of the semantics of a system (by introducing strongly consistent operations to an eventually consistent system) results in the weakening of the guarantees on the eventually consistent operations.

cs.DC↗

arXiv2019

Creek: Low-latency, Mixed-Consistency Transactional Replication Scheme

In this paper we introduce Creek, a low-latency, eventually consistent replication scheme that also enables execution of strongly consistent operations (akin to ACID transactions). Operations can have arbitrary complex (but deterministic) semantics. Similarly to state machine replication (SMR), Creek totally-orders all operations, but does so using two different broadcast mechanisms: a timestamp-based one and our novel conditional atomic broadcast (CAB). The former is used to establish a tentative order of all operations for speculative execution, and it can tolerate network partitions. On the other hand, CAB is only used to ensure linearizable execution of the strongly consistent operations, whenever distributed consensus can be solved. The execution of strongly consistent operations also stabilizes the execution order of the causally related weakly consistent operations. Creek uses multiversion concurrency control to efficiently handle operations' rollbacks and reexecutions resulting from the mismatch between the tentative and the final execution orders. In the TPC-C benchmark, Creek offers up to 2.5 times lower latency in returning client responses compared to the state-of-the-art speculative SMR scheme, while maintaining high accuracy of the speculative execution (92-100%).

cs.DC↗

arXiv2016

Hybrid Transactional Replication: State-Machine and Deferred-Update Replication Combined

We propose Hybrid Transactional Replication (HTR), a novel replication scheme for highly dependable services. It combines two schemes: a transaction is executed either optimistically by only one service replica in the deferred update mode (DU), or deterministically by all replicas in the state machine mode (SM); the choice is made by an oracle. The DU mode allows for parallelism and thus takes advantage of multicore hardware. In contrast to DU, the SM mode guarantees abort-free execution, so it is suitable for irrevocable operations and transactions generating high contention. For expressiveness, transactions can be discarded or retried on demand. We formally prove that the higher flexibility of the scheme does not come at the cost of weaker guarantees for clients: HTR satisfies strong consistency guarantees akin to those provided by other popular transactional replication schemes such as Deferred Update Replication. We developed HTR-enabled Paxos STM, an object-based distributed transactional memory system, and evaluated it thoroughly under various workloads. We show the benefits of using a novel oracle, which relies on machine learning techniques for automatic adaptation to changing conditions. In our tests, the ML-based oracle provides up to 50% improvement in throughput when compared to the system running with DU-only or SM-only oracles. Our approach is inspired by a well known algorithm used in the context of the multi-armed bandit problem.

cs.DC↗