Mark Silberstein
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Mark Silberstein has a total of 71 co-authors. Here are the top co-authors by number of shared publications:
- Alon Rashelbach
Technion, Haifa, Israel
6 shared publications - Ori Rottenstreich
Technion, Haifa, Israel
4 shared publications - Shai Bergman
Huawei, Zurich, Switzerland
3 shared publications - Ang Chen
University of Michigan, Ann Arbor, USA
2 shared publications - Haggai Eran
Nvidia Corp
2 shared publications - Oleksii Oleksenko
Azure Research, Microsoft, Cambridge, United Kingdom
2 shared publications
Latest Publications
- Open Access
- 24 June 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Computer Systems
- Vol. 44 (3) , 1-35
FlexDriver: A Network Driver for Your Accelerator
We propose a new system design for connecting hardware and FPGA accelerators to the network, allowing them to directly control commodity ASIC NICs without using the CPU. This solves the key challenge of leveraging existing NIC hardware offloads such as RDMA and virtualization for hardware disaggregation and accelerator networking. Our approach supports a diverse set of use cases, from direct network access for disaggregated accelerators to inline acceleration of the network stack and disaggregation of main system memory, all without implementing complex networking logic. To demonstrate this approach, we build FlexDriver (FLD) , a hardware module that implements a NIC data-plane driver. Our main technical contribution is compressing NIC control structures by 5 ×, allowing FLD to achieve high scalability with low die area and no memory bandwidth interference. We build two prototypes – FLD core on NVIDIA Innova-2 FPGA SmartNICs and FLD with a load/store interface on IBM OpenCAPI FPGA deployment with ConnectX-6 Dx. We demonstrate four different use cases: a disaggregated LTE cipher, an IP-reassembly inline accelerator, an IoT cryptographic-token authentication offload, and a fine-grained memory disaggregation datapath over RDMA. These leverage the ASIC NIC for RDMA processing, VXLAN tunneling, and traffic shaping, without CPU involvement.
- 24 June 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Computer Systems
- Vol. 44 (3) , 1-35
Open AccessFlexDriver: A Network Driver for Your Accelerator
We propose a new system design for connecting hardware and FPGA accelerators to the network, allowing them to directly control commodity ASIC NICs without using the CPU. This solves the key challenge of leveraging existing NIC hardware offloads such as RDMA and virtualization for hardware disaggregation and accelerator networking. Our approach supports a diverse set of use cases, from direct network access for disaggregated accelerators to inline acceleration of the network stack and disaggregation of main system memory, all without implementing complex networking logic. To demonstrate this approach, we build FlexDriver (FLD) , a hardware module that implements a NIC data-plane driver. Our main technical contribution is compressing NIC control structures by 5 ×, allowing FLD to achieve high scalability with low die area and no memory bandwidth interference. We build two prototypes – FLD core on NVIDIA Innova-2 FPGA SmartNICs and FLD with a load/store interface on IBM OpenCAPI FPGA deployment with ConnectX-6 Dx. We demonstrate four different use cases: a disaggregated LTE cipher, an IP-reassembly inline accelerator, an IoT cryptographic-token authentication offload, and a fine-grained memory disaggregation datapath over RDMA. These leverage the ASIC NIC for RDMA processing, VXLAN tunneling, and traffic shaping, without CPU involvement.
- Open Access
- 17 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 139-148
FlowPulse: Catching Network Failures in ML Clusters
- 17 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 139-148
Open AccessFlowPulse: Catching Network Failures in ML Clusters
- Open Access
- 6 November 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Computer Systems
- Vol. 43 (4) , 1-2
Introduction to the Special Section on EuroSys 2024
- 6 November 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Computer Systems
- Vol. 43 (4) , 1-2
Open AccessIntroduction to the Special Section on EuroSys 2024
- Open Access
- 8 September 2025
- Published by Association for Computing Machinery (ACM)
- p. 19-25
NSX: Large-Scale Network Simulation on an AI Server
- 8 September 2025
- Published by Association for Computing Machinery (ACM)
- p. 19-25
Open AccessNSX: Large-Scale Network Simulation on an AI Server
- Open Access
- 30 March 2025
- Published by Association for Computing Machinery (ACM)
- p. 32-47
AMuLeT: Automated Design-Time Testing of Secure Speculation Countermeasures
- 30 March 2025
- Published by Association for Computing Machinery (ACM)
- p. 32-47
Open AccessAMuLeT: Automated Design-Time Testing of Secure Speculation Countermeasures
- Open Access
- 16 September 2024
- Published by Association for Computing Machinery (ACM)
- p. 146-161
Space-efficient FTL for Mobile Storage via Tiny Neural Nets
We present RQFTL, a demand-based FTL for mobile storage controllers that boosts the effective Logical-To-Physical (L2P) address translation cache capacity over state-of-the-art techniques. RQFTL stores a large part of the L2P cache in a compressed form, and employs a learned data structure called RQRMI that leverages tiny neural nets to quickly find the correct translation entry in the cache. RQFTL uses neural network inference for cache lookups, and rapidly retrains the neural nets to efficiently handle L2P cache updates. It is specifically optimized to achieve high coverage for scattered read accesses, making it suitable for popular read-skewed workloads such as mobile gaming. We evaluate RQFTL on hours-long real-world I/O traces of popular modern mobile apps, including games, video editing, and social networking apps collected on Google Pixel 6a phone. We show that RQFTL outperforms all the state-of-the-art FTLs in these workloads, increasing the effective L2P cache capacity by over an order of magnitude compared to DFTL and up to 5x over the recent LeaFTL. As a result, it achieves 65%, and 25% lower miss rate compared to DFTL and LeaFTL respectively, under the same SRAM capacity, and allows reduction of the total SRAM capacity of a controller by about a third of that of LeaFTL.
- 16 September 2024
- Published by Association for Computing Machinery (ACM)
- p. 146-161
Open AccessSpace-efficient FTL for Mobile Storage via Tiny Neural Nets
We present RQFTL, a demand-based FTL for mobile storage controllers that boosts the effective Logical-To-Physical (L2P) address translation cache capacity over state-of-the-art techniques. RQFTL stores a large part of the L2P cache in a compressed form, and employs a learned data structure called RQRMI that leverages tiny neural nets to quickly find the correct translation entry in the cache. RQFTL uses neural network inference for cache lookups, and rapidly retrains the neural nets to efficiently handle L2P cache updates. It is specifically optimized to achieve high coverage for scattered read accesses, making it suitable for popular read-skewed workloads such as mobile gaming. We evaluate RQFTL on hours-long real-world I/O traces of popular modern mobile apps, including games, video editing, and social networking apps collected on Google Pixel 6a phone. We show that RQFTL outperforms all the state-of-the-art FTLs in these workloads, increasing the effective L2P cache capacity by over an order of magnitude compared to DFTL and up to 5x over the recent LeaFTL. As a result, it achieves 65%, and 25% lower miss rate compared to DFTL and LeaFTL respectively, under the same SRAM capacity, and allows reduction of the total SRAM capacity of a controller by about a third of that of LeaFTL.
- Open Access
- 4 August 2024
- Published by Association for Computing Machinery (ACM)
- p. 735-749
In-Network Address Caching for Virtual Networks
Packet routing in virtual networks requires virtual-to-physical address translation. The address mappings are updated by a single party, i.e., the network administrator, but they are read by multiple devices across the network when routing tenant packets. Existing approaches face an inherent read-write performance tradeoff: they either store these mappings in dedicated gateways for fast updates at the cost of slower forwarding or replicate them at end-hosts and suffer from slow updates. SwitchV2P aims to escape this tradeoff by leveraging the network switches to transparently cache the address mappings while learning them from the traffic. SwitchV2P brings the mappings closer to the sender, thus reducing the first packet latency and translation overheads, while simultaneously enabling fast mapping updates, all without changing existing routing policies and deployed gateways. The topology-aware data-plane caching protocol allows the switches to transparently adapt to changing network conditions and varying in-switch memory capacity. Our evaluation shows the benefits of in-network address mapping, including an up to 7.8× and 4.3× reduction in FCT and first packet latency respectively, and a substantial reduction in translation gateway load. Additionally, SwitchV2P achieves up to a 1.9× reduction in bandwidth overheads and requires order-of-magnitude fewer gateways for equivalent performance.
- 4 August 2024
- Published by Association for Computing Machinery (ACM)
- p. 735-749
Open AccessIn-Network Address Caching for Virtual Networks
Packet routing in virtual networks requires virtual-to-physical address translation. The address mappings are updated by a single party, i.e., the network administrator, but they are read by multiple devices across the network when routing tenant packets. Existing approaches face an inherent read-write performance tradeoff: they either store these mappings in dedicated gateways for fast updates at the cost of slower forwarding or replicate them at end-hosts and suffer from slow updates. SwitchV2P aims to escape this tradeoff by leveraging the network switches to transparently cache the address mappings while learning them from the traffic. SwitchV2P brings the mappings closer to the sender, thus reducing the first packet latency and translation overheads, while simultaneously enabling fast mapping updates, all without changing existing routing policies and deployed gateways. The topology-aware data-plane caching protocol allows the switches to transparently adapt to changing network conditions and varying in-switch memory capacity. Our evaluation shows the benefits of in-network address mapping, including an up to 7.8× and 4.3× reduction in FCT and first packet latency respectively, and a substantial reduction in translation gateway load. Additionally, SwitchV2P achieves up to a 1.9× reduction in bandwidth overheads and requires order-of-magnitude fewer gateways for equivalent performance.
- Open Access
- 14 June 2024
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 121 (25)
Algorithm-assisted discovery of an intrinsic order among mathematical constants
In recent decades, a growing number of discoveries in mathematics have been assisted by computer algorithms, primarily for exploring large parameter spaces. As computers become more powerful, an intriguing possibility arises—the interplay between human intuition and computer algorithms can lead to discoveries of mathematical structures that would otherwise remain elusive. Here, we demonstrate computer-assisted discovery of a previously unknown mathematical structure, the conservative matrix field. In the spirit of the Ramanujan Machine project, we developed a massively parallel computer algorithm that found a large number of formulas, in the form of continued fractions, for numerous mathematical constants. The patterns arising from those formulas enabled the construction of the first conservative matrix fields and revealed their overarching properties. Conservative matrix fields unveil unexpected relations between different mathematical constants, such as π and ln(2), or e and the Gompertz constant. The importance of these matrix fields is further realized by their ability to connect formulas that do not have any apparent relation, thus unifying hundreds of existing formulas and generating infinitely many new formulas. We exemplify these implications on values of the Riemann zeta function ζ (n), studied for centuries across mathematics and physics. Matrix fields also enable new mathematical proofs of irrationality. For example, we use them to generalize the celebrated proof by Apéry of the irrationality of ζ (3). Utilizing thousands of personal computers worldwide, our research strategy demonstrates the power of large-scale computational approaches to tackle longstanding open problems and discover unexpected connections across diverse fields of science.
- 14 June 2024
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 121 (25)
Open AccessAlgorithm-assisted discovery of an intrinsic order among mathematical constants
In recent decades, a growing number of discoveries in mathematics have been assisted by computer algorithms, primarily for exploring large parameter spaces. As computers become more powerful, an intriguing possibility arises—the interplay between human intuition and computer algorithms can lead to discoveries of mathematical structures that would otherwise remain elusive. Here, we demonstrate computer-assisted discovery of a previously unknown mathematical structure, the conservative matrix field. In the spirit of the Ramanujan Machine project, we developed a massively parallel computer algorithm that found a large number of formulas, in the form of continued fractions, for numerous mathematical constants. The patterns arising from those formulas enabled the construction of the first conservative matrix fields and revealed their overarching properties. Conservative matrix fields unveil unexpected relations between different mathematical constants, such as π and ln(2), or e and the Gompertz constant. The importance of these matrix fields is further realized by their ability to connect formulas that do not have any apparent relation, thus unifying hundreds of existing formulas and generating infinitely many new formulas. We exemplify these implications on values of the Riemann zeta function ζ (n), studied for centuries across mathematics and physics. Matrix fields also enable new mathematical proofs of irrationality. For example, we use them to generalize the celebrated proof by Apéry of the irrationality of ζ (3). Utilizing thousands of personal computers worldwide, our research strategy demonstrates the power of large-scale computational approaches to tackle longstanding open problems and discover unexpected connections across diverse fields of science.
- Open Access
- 28 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 886-899
NeuroLPM - Scaling Longest Prefix Match Hardware with Neural Networks
Longest Prefix Match engines (LPM) are broadly used in computer systems and especially in modern network devices such as Network Interface Cards (NICs), switches and routers. However, existing LPM hardware fails to scale to millions of rules required by modern systems, is often optimized for specific applications, and thus is performance-sensitive to the structure of LPM rules. We describe NeuroLPM, a new architecture for multi-purpose LPM hardware that replaces queries in traditional memory-intensive trie- and hash-table data structures with inference in a lightweight Neural Network-based model, called RQRMI. NeuroLPM scales to millions of rules under small on-die SRAM budget and achieves stable, rule-structure-agnostic performance, allowing its use in a variety of applications. We solve several unique challenges when implementing RQRMI inference in hardware, including minimizing the amount of floating point computations while maintaining query correctness, and scaling the rule-set size while ensuring small, deterministic off-chip memory bandwidth. We prototype NeuroLPM in Verilog and evaluate it on real-world packet forwarding rule-sets and network traces. NeuroLPM offers substantial scalability benefits without any application-specific optimizations. For example, it is the only algorithm that can serve a 950K-large rule-set at an average of 196M queries per second with 4.5MB of SRAM, only within 2% of the best-case throughput of the state-of-the-art Tree Bitmap and SAIL on smaller rule-sets. With 2MB of SRAM, it reduces the DRAM bandwidth per query, the dominant performance factor, by up to 9 × and 3 × compared to the state-of-the-art.
- 28 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 886-899
Open AccessNeuroLPM - Scaling Longest Prefix Match Hardware with Neural Networks
Longest Prefix Match engines (LPM) are broadly used in computer systems and especially in modern network devices such as Network Interface Cards (NICs), switches and routers. However, existing LPM hardware fails to scale to millions of rules required by modern systems, is often optimized for specific applications, and thus is performance-sensitive to the structure of LPM rules. We describe NeuroLPM, a new architecture for multi-purpose LPM hardware that replaces queries in traditional memory-intensive trie- and hash-table data structures with inference in a lightweight Neural Network-based model, called RQRMI. NeuroLPM scales to millions of rules under small on-die SRAM budget and achieves stable, rule-structure-agnostic performance, allowing its use in a variety of applications. We solve several unique challenges when implementing RQRMI inference in hardware, including minimizing the amount of floating point computations while maintaining query correctness, and scaling the rule-set size while ensuring small, deterministic off-chip memory bandwidth. We prototype NeuroLPM in Verilog and evaluate it on real-world packet forwarding rule-sets and network traces. NeuroLPM offers substantial scalability benefits without any application-specific optimizations. For example, it is the only algorithm that can serve a 950K-large rule-set at an average of 196M queries per second with 4.5MB of SRAM, only within 2% of the best-case throughput of the state-of-the-art Tree Bitmap and SAIL on smaller rule-sets. With 2MB of SRAM, it reduces the DRAM bandwidth per query, the dominant performance factor, by up to 9 × and 3 × compared to the state-of-the-art.
- Open Access
- 3 September 2023
- Published by Association for Computing Machinery (ACM)
- p. 1-2
Nucleotide String Indexing using Range Matching
- 3 September 2023
- Published by Association for Computing Machinery (ACM)
- p. 1-2
Open AccessNucleotide String Indexing using Range Matching