Reza Salkhordeh
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Reza Salkhordeh has a total of 91 co-authors. Here are the top co-authors by number of shared publications:
- Hossein Asadi
Sharif University of Technology
16 shared publications - André Brinkmann
Saarland University, Saarbrücken, Germany
14 shared publications - Shahriar Ebrahimi
Department of Computer Engineering, Sharif University of Technology, Tehran, Iran
4 shared publications - Ali Sedaghatgoo
Sharif University of Technology, Tehran, Iran
3 shared publications - Eric Borba
Johannes Gutenberg-Universität Mainz, Mainz, Germany
3 shared publications - Nafiseh Moti
Johannes Gutenberg University Mainz, Germany
3 shared publications
Latest Publications
- Open Access
- 20 July 2026
- Published by Association for Computing Machinery (ACM) in ACM SIGMETRICS Performance Evaluation Review
- Vol. 54 (1) , 123-125
Towards Scalable Storage Architectures for GPU Clusters Running Large Language Models
Accelerators in multi-GPU nodes increasingly sit idle due to storage and I/O inefficiencies. We present a phase-aware evaluation of storage datapaths for Large Language Models (LLMs) on single-node, multi-GPU systems, comparing libaio, io_uring, Storage Performance Development Kit (SPDK), and GPUDirect Storage (GDS) across SATA SSDs, NVMe SSDs, Optane NVMe, and Optane Persistent Memory. Using an automated framework, we test over 25,000 configurations and measure throughput, latency, Input Output Per Second (IOPS), and CPU overhead using workloads derived from real LLM traces spanning pre-training, fine-tuning, and inference. For inference workloads dominated by small random I/O, io_uring provides the best latency and competitive IOPS on NVMe devices. For pre-training and fine-tuning, which rely on large sequential transfers, GDS significantly reduces model load times and CPU usage. These findings show that optimal storage datapaths are phase-dependent and offer practical guidance for narrowing the storage--compute gap in GPU-based LLM systems.
- 20 July 2026
- Published by Association for Computing Machinery (ACM) in ACM SIGMETRICS Performance Evaluation Review
- Vol. 54 (1) , 123-125
Open AccessTowards Scalable Storage Architectures for GPU Clusters Running Large Language Models
Accelerators in multi-GPU nodes increasingly sit idle due to storage and I/O inefficiencies. We present a phase-aware evaluation of storage datapaths for Large Language Models (LLMs) on single-node, multi-GPU systems, comparing libaio, io_uring, Storage Performance Development Kit (SPDK), and GPUDirect Storage (GDS) across SATA SSDs, NVMe SSDs, Optane NVMe, and Optane Persistent Memory. Using an automated framework, we test over 25,000 configurations and measure throughput, latency, Input Output Per Second (IOPS), and CPU overhead using workloads derived from real LLM traces spanning pre-training, fine-tuning, and inference. For inference workloads dominated by small random I/O, io_uring provides the best latency and competitive IOPS on NVMe devices. For pre-training and fine-tuning, which rely on large sequential transfers, GDS significantly reduces model load times and CPU usage. These findings show that optimal storage datapaths are phase-dependent and offer practical guidance for narrowing the storage--compute gap in GPU-based LLM systems.
- Open Access
- 8 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 123-125
Towards Scalable Storage Architectures for GPU Clusters Running Large Language Models
- 8 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 123-125
Open AccessTowards Scalable Storage Architectures for GPU Clusters Running Large Language Models
- Open Access
- 26 March 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Measurement and Analysis of Computing Systems
- Vol. 10 (1) , 1-26
Towards Scalable Storage Architectures for GPU Clusters Running Large Language Models
In modern multi-accelerator nodes, GPU throughput is increasingly constrained by storage and I/O bottlenecks, leaving accelerators idle as data transfer is restricted by software. In this study, we focus on single-node, multi-GPU systems with small-to-medium-scale models and present a systematic, phase-aware evaluation of datapath performance for Large Language Models (LLMs), covering technologies such as in-kernel libaio, hybrid user-kernel io_uring, user-space NVMe via the Storage Performance Development Kit (SPDK), and GPUDirect Storage (GDS). These approaches are evaluated across various storage media including SATA Solid State Drives (SSDs), NVMe SSDs, Optane NVMe, and Optane Persistent Memory (PMem). Leveraging an automated evaluation framework, we explore over 25,000 configurations, measuring throughput, latency, I/O per second (IOPS), and CPU cost. Our study offers LLM storage scenarios in both standardized benchmarks and real-world production traces, ensuring that our workload models accurately reflect the I/O demands across pre-training, fine-tuning, and inference. We find that for inference, io_uring achieves the lowest latency and competitive IOPS for small random I/O on NVMe. In contrast, SPDK is limited to raw block-device evaluation due to its lack of POSIX file-system support. For pre-training and fine-tuning, workloads are dominated by coarse-grained sequential reads and writes, where GDS excels in reducing load times and host CPU usage. Among CPU-mediated datapaths, CPU efficiency—measured as GB/s per core—emerges as the key differentiator. Taken together, these results yield actionable design guidelines: align the choice of datapath with the LLM pipeline phase. Use io_uring for inference to optimize data transfer efficiency and minimize latency, and leverage GDS for pre-training and fine-tuning to improve throughput per core, thereby narrowing the storage-to-compute gap in GPU LLM clusters.
- 26 March 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Measurement and Analysis of Computing Systems
- Vol. 10 (1) , 1-26
Open AccessTowards Scalable Storage Architectures for GPU Clusters Running Large Language Models
In modern multi-accelerator nodes, GPU throughput is increasingly constrained by storage and I/O bottlenecks, leaving accelerators idle as data transfer is restricted by software. In this study, we focus on single-node, multi-GPU systems with small-to-medium-scale models and present a systematic, phase-aware evaluation of datapath performance for Large Language Models (LLMs), covering technologies such as in-kernel libaio, hybrid user-kernel io_uring, user-space NVMe via the Storage Performance Development Kit (SPDK), and GPUDirect Storage (GDS). These approaches are evaluated across various storage media including SATA Solid State Drives (SSDs), NVMe SSDs, Optane NVMe, and Optane Persistent Memory (PMem). Leveraging an automated evaluation framework, we explore over 25,000 configurations, measuring throughput, latency, I/O per second (IOPS), and CPU cost. Our study offers LLM storage scenarios in both standardized benchmarks and real-world production traces, ensuring that our workload models accurately reflect the I/O demands across pre-training, fine-tuning, and inference. We find that for inference, io_uring achieves the lowest latency and competitive IOPS for small random I/O on NVMe. In contrast, SPDK is limited to raw block-device evaluation due to its lack of POSIX file-system support. For pre-training and fine-tuning, workloads are dominated by coarse-grained sequential reads and writes, where GDS excels in reducing load times and host CPU usage. Among CPU-mediated datapaths, CPU efficiency—measured as GB/s per core—emerges as the key differentiator. Taken together, these results yield actionable design guidelines: align the choice of datapath with the LLM pipeline phase. Use io_uring for inference to optimize data transfer efficiency and minimize latency, and leverage GDS for pre-training and fine-tuning to improve throughput per core, thereby narrowing the storage-to-compute gap in GPU LLM clusters.
- Open Access
- 3 March 2026
- Published by Elsevier in Journal of Systems Architecture
- Vol. 175
ZTL: A block layer ZNS driver
- 3 March 2026
- Published by Elsevier in Journal of Systems Architecture
- Vol. 175
Open AccessZTL: A block layer ZNS driver
- 23 October 2025
- Published by Springer Nature
- p. 125-139
ZTL: Enabling Zoned Namespace Support for File Systems
- 23 October 2025
- Published by Springer Nature
- p. 125-139
ZTL: Enabling Zoned Namespace Support for File Systems
- 24 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. PP (10459219) , 1-16
ELICA: Efficient and Load Balanced I/O Cache Architecture for Hyperconverged Infrastructures
Hyperconverged Infrastructures (HCIs) combine processing and storage elements to meet the requirements of data-intensive applications in performance, scalability, and quality of service. As an emerging paradigm, HCI should couple with a variety of traditional performance improvement approaches such as I/O caching in virtualized platforms. Contemporary I/O caching schemes are optimized for traditional single-node storage architectures and suffer from two major shortcomings for multi-node architectures: a) imbalanced cache space requirement and b) imbalanced I/O traffic and load. This makes existing schemes inefficient in distributing cache resources over an array of separate physical nodes. In this paper, we propose an Efficient and Load Balanced I/O Cache Architecture (ELICA), managing the solid-state drive (SSD) cache resources across HCI nodes to enhance I/O performance. ELICA dynamically reconfigures and distributes the SSD cache resources throughout the array of HCI nodes and also balances the network traffic and I/O cache load by dynamic reallocation of cache resources. To maximize the performance, we further present an optimization problem defined by Integer Linear Programming to efficiently distribute cache resources and balance the network traffic and I/O cache relocations. Our experimental results on a real platform show that ELICA improves quality of service in terms of average and worst-case latency in HCIs by 3.1× and 23%, respectively, compared to the state-of-the-art.
- 24 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. PP (10459219) , 1-16
ELICA: Efficient and Load Balanced I/O Cache Architecture for Hyperconverged Infrastructures
Hyperconverged Infrastructures (HCIs) combine processing and storage elements to meet the requirements of data-intensive applications in performance, scalability, and quality of service. As an emerging paradigm, HCI should couple with a variety of traditional performance improvement approaches such as I/O caching in virtualized platforms. Contemporary I/O caching schemes are optimized for traditional single-node storage architectures and suffer from two major shortcomings for multi-node architectures: a) imbalanced cache space requirement and b) imbalanced I/O traffic and load. This makes existing schemes inefficient in distributing cache resources over an array of separate physical nodes. In this paper, we propose an Efficient and Load Balanced I/O Cache Architecture (ELICA), managing the solid-state drive (SSD) cache resources across HCI nodes to enhance I/O performance. ELICA dynamically reconfigures and distributes the SSD cache resources throughout the array of HCI nodes and also balances the network traffic and I/O cache load by dynamic reallocation of cache resources. To maximize the performance, we further present an optimization problem defined by Integer Linear Programming to efficiently distribute cache resources and balance the network traffic and I/O cache relocations. Our experimental results on a real platform show that ELICA improves quality of service in terms of average and worst-case latency in HCIs by 3.1× and 23%, respectively, compared to the state-of-the-art.
- Open Access
- 8 February 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Storage
- Vol. 21 (2) , 1-27
HLN-Tree: A memory-efficient B+-Tree with huge leaf nodes and locality predictors
Key-value stores in Cloud environments can contain more than 2 45 unique elements and be larger than 100 PByte. B + -Trees are well suited for these larger-than-memory datasets and seamlessly index data stored on thousands of secondary storage devices. Unfortunately, it is often uneconomical to even store all inner tree nodes in memory for these dataset sizes. Therefore, lookup performance is affected by the additional IOs for reading inner nodes. This number of inner nodes can be reduced by increasing the size of leaf nodes. We propose HLN-Trees, which support huge leaf nodes without increasing the IO sizes for individual index operations. They partition leaf nodes in arrays of independent subnodes and combine ideas from BD-trees with rebalancing, learning key deviations, and storing locality predictors. HLN-Trees have been initially designed for uniform random key distributions and support arbitrary key distributions through an additional layer of hashing in leaf nodes. HLN-Trees decrease the number of inner nodes by up to 256× for uniform random key distributions and by 16× to 64× for arbitrary ones compared to B + -Trees, while keeping their performance at the same level even at high concurrency levels. We show analytically and through real-world and synthetic benchmarks that HLN-Trees also outperform state-of-the-art learned indexes for secondary storage.
- 8 February 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Storage
- Vol. 21 (2) , 1-27
Open AccessHLN-Tree: A memory-efficient B+-Tree with huge leaf nodes and locality predictors
Key-value stores in Cloud environments can contain more than 2 45 unique elements and be larger than 100 PByte. B + -Trees are well suited for these larger-than-memory datasets and seamlessly index data stored on thousands of secondary storage devices. Unfortunately, it is often uneconomical to even store all inner tree nodes in memory for these dataset sizes. Therefore, lookup performance is affected by the additional IOs for reading inner nodes. This number of inner nodes can be reduced by increasing the size of leaf nodes. We propose HLN-Trees, which support huge leaf nodes without increasing the IO sizes for individual index operations. They partition leaf nodes in arrays of independent subnodes and combine ideas from BD-trees with rebalancing, learning key deviations, and storing locality predictors. HLN-Trees have been initially designed for uniform random key distributions and support arbitrary key distributions through an additional layer of hashing in leaf nodes. HLN-Trees decrease the number of inner nodes by up to 256× for uniform random key distributions and by 16× to 64× for arbitrary ones compared to B + -Trees, while keeping their performance at the same level even at high concurrency levels. We show analytically and through real-world and synthetic benchmarks that HLN-Trees also outperform state-of-the-art learned indexes for secondary storage.
- Open Access
- 6 August 2024
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Storage
- Vol. 20 (4) , 1-28
From SSDs Back to HDDs: Optimizing VDO to Support Inline Deduplication and Compression for HDDs as Primary Storage Media
Deduplication and compression are powerful techniques to reduce the ratio between the quantity of logical data stored and the physical amount of consumed storage. Deduplication can impose significant performance overheads, as duplicate detection for large systems induces random accesses to the backend storage. These random accesses have led to the concern that deduplication for primary storage and HDDs are not compatible. Most inline data reduction solutions are therefore optimized for SSDs and discourage their use for HDDs, even for sequential workloads. In this work, we show that these concerns are valid if and only if the lessons learned from deduplication research are not applied. We have therefore investigated data reduction solutions for primary storage based on the RedHat Virtual Disk Optimizer (VDO) and show that directly applying them can decrease sequential write performance for HDDs by 36×. We then show that slight modifications to VDO plus the integration of a very small SSD area significantly improve performance even beyond the performance without data reduction enabled, making HDDs more cost-efficient for a wide range of mostly sequential cloud workloads than SSDs. Additionally, these VDO optimizations do not require to maintain different code bases for HDDs and SSDs, and we therefore provide the first data reduction solution applicable to both storage media.
- 6 August 2024
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Storage
- Vol. 20 (4) , 1-28
Open AccessFrom SSDs Back to HDDs: Optimizing VDO to Support Inline Deduplication and Compression for HDDs as Primary Storage Media
Deduplication and compression are powerful techniques to reduce the ratio between the quantity of logical data stored and the physical amount of consumed storage. Deduplication can impose significant performance overheads, as duplicate detection for large systems induces random accesses to the backend storage. These random accesses have led to the concern that deduplication for primary storage and HDDs are not compatible. Most inline data reduction solutions are therefore optimized for SSDs and discourage their use for HDDs, even for sequential workloads. In this work, we show that these concerns are valid if and only if the lessons learned from deduplication research are not applied. We have therefore investigated data reduction solutions for primary storage based on the RedHat Virtual Disk Optimizer (VDO) and show that directly applying them can decrease sequential write performance for HDDs by 36×. We then show that slight modifications to VDO plus the integration of a very small SSD area significantly improve performance even beyond the performance without data reduction enabled, making HDDs more cost-efficient for a wide range of mostly sequential cloud workloads than SSDs. Additionally, these VDO optimizations do not require to maintain different code bases for HDDs and SSDs, and we therefore provide the first data reduction solution applicable to both storage media.
- 1 August 2024
- Published by Springer Nature
- p. 266-281
A Hierarchical Modeling Approach for Assessing the Reliability and Performability of Burst Buffers
- 1 August 2024
- Published by Springer Nature
- p. 266-281
A Hierarchical Modeling Approach for Assessing the Reliability and Performability of Burst Buffers
- 19 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. 35 (12) , 2608-2623
HybRAID: A High-Performance Hybrid RAID Storage Architecture for Write-Intensive Applications in All-Flash Storage Systems
With the ever-increasing demand for higher I/O performance and reliability in data-intensive applications, solid-state drives (SSDs) typically configured as redundant array of independent disks (RAID) are broadly used in enterprise all-flash storage systems . While a mirrored RAID offers higher performance in random access workloads, parity-based RAIDs (e.g., RAID5) provide higher performance in sequential accesses with less cost overhead. Previous studies try to address the poor performance of parity-based RAIDs in small writes (i.e., writes into a single disk) by offering various schemes, including caching or logging small writes. However, such techniques impose a significant performance and/or reliability overheads and are seldom used in the industry. In addition, our empirical analysis shows that partial stripe writes, i.e., writing into a fraction of a full array in parity-based RAIDs, can significantly degrade the I/O performance, which has not been addressed in the previous work. In this paper, we first offer an empirical study which reveals partial stripe writes reduce the performance of parity-based RAIDs by up to 6.85× compared to full stripe writes (i.e., writes into entire disks). Then, we propose a high-performance hyb rid RAID storage architecture, called HybRAID , which is optimized for write-intensive applications. HybRAID exploits the advantages of mirror- and parity-based RAIDs to improve the write performance. HybRAID directs a) aligned full stripe writes to parity-based RAID tier and b) small/partial stripe writes to the RAID1 tier. We propose an online migration scheme, which aims to move small/partial writes from parity-based RAID to RAID1, based on access frequency of updates. As a complement, we further offer offline migration, whose aim is to make room in the fast tier for future references. Experimental results over enterprise SSDs show that HybRAID improves the performance of write-intensive applications by 3.3× and 2.6×, as well as enhancing performance per cost by 3.1× and 3.0× compared to parity-based RAID and RAID10, respectively, at equivalent costs.
- 19 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. 35 (12) , 2608-2623
HybRAID: A High-Performance Hybrid RAID Storage Architecture for Write-Intensive Applications in All-Flash Storage Systems
With the ever-increasing demand for higher I/O performance and reliability in data-intensive applications, solid-state drives (SSDs) typically configured as redundant array of independent disks (RAID) are broadly used in enterprise all-flash storage systems . While a mirrored RAID offers higher performance in random access workloads, parity-based RAIDs (e.g., RAID5) provide higher performance in sequential accesses with less cost overhead. Previous studies try to address the poor performance of parity-based RAIDs in small writes (i.e., writes into a single disk) by offering various schemes, including caching or logging small writes. However, such techniques impose a significant performance and/or reliability overheads and are seldom used in the industry. In addition, our empirical analysis shows that partial stripe writes, i.e., writing into a fraction of a full array in parity-based RAIDs, can significantly degrade the I/O performance, which has not been addressed in the previous work. In this paper, we first offer an empirical study which reveals partial stripe writes reduce the performance of parity-based RAIDs by up to 6.85× compared to full stripe writes (i.e., writes into entire disks). Then, we propose a high-performance hyb rid RAID storage architecture, called HybRAID , which is optimized for write-intensive applications. HybRAID exploits the advantages of mirror- and parity-based RAIDs to improve the write performance. HybRAID directs a) aligned full stripe writes to parity-based RAID tier and b) small/partial stripe writes to the RAID1 tier. We propose an online migration scheme, which aims to move small/partial writes from parity-based RAID to RAID1, based on access frequency of updates. As a complement, we further offer offline migration, whose aim is to make room in the fast tier for future references. Experimental results over enterprise SSDs show that HybRAID improves the performance of write-intensive applications by 3.3× and 2.6×, as well as enhancing performance per cost by 3.1× and 3.0× compared to parity-based RAID and RAID10, respectively, at equivalent costs.