DapuStor J5060 Specifications

DapuStor J5060 Performance

Checkpointing

To evaluate the Dapustor J5060 SSD’s real-world performance in AI training environments, we utilized the Data and Learning Input/Output (DLIO) benchmark tool. Developed by Argonne National Laboratory, DLIO is specifically designed to test I/O patterns in deep learning workloads. It provides insights into how storage systems handle challenges such as checkpointing, data ingestion, and model training. The chart below illustrates how both drives handle the process across 99 checkpoints. When training machine learning models, checkpoints are essential for saving the model’s state periodically, preventing loss of progress during interruptions or power failures. This storage demand requires robust performance, especially under sustained or intensive workloads.

The platform chosen for this work was our Dell PowerEdge R760 running Ubuntu 22.04.02 LTS. We used DLIO benchmark version 2.0 from the August 13, 2024, release. Our system configuration is outlined below:

• 2 x Intel Xeon Gold 6430 (32-Core, 2.1GHz)
• 16 x 64GB DDR5-4400
• 480GB Dell BOSS SSD
• Serial Cables Gen5 JBOF
  • 61.44TB Dapustor J5060
  • 61.44TB Solidigm D5-P5336

To ensure our benchmarking reflected real-world scenarios, we based our testing on the LLAMA 3.1 405B model architecture. We implemented checkpointing using torch.save() to capture model parameters, optimizer states, and layer states. Our setup simulated an 8-GPU system, implementing a hybrid parallelism strategy with 4-way tensor parallel and 2-way pipeline parallel processing distributed across the eight GPUs. This configuration resulted in checkpoint sizes of 1,636GB, representative of modern large language model training requirements.

Overall, the Dapustor J5060 demonstrated solid consistency during the initial phase of testing, with times hovering around 575.66 seconds for the first 33 checkpoints. The 5060J was able to maintain higher performance before the drive was filled for the first time. On the other hand, the Solidigm P5336, while initially slower than the J5060, demonstrated consistent performance as testing continued.

When considering the overall averages, the Dapustor J5060 posted a time of 769.44 seconds, while the Solidigm P5336 finished in 640.17 seconds. This puts the Solidigm P5336 ahead in terms of saving checkpoints faster.

Overall, the Dapustor J5060 handles shorter operations well but struggles with sustained writes beyond 30 minutes. Meanwhile, the Solidigm P5336 is the better drive for consistent performance throughout prolonged tasks. This weaker write performance from the Dapustor J5060 is evident when its checkpointing speed deteriorates as the test continues.

GPU Direct Storage

GPU Direct Storage is a technology that enables direct data transfer between storage devices and GPUs, bypassing the CPU and system memory. In traditional data transfer, data is read from storage into the CPU’s memory and then copied to the GPU’s memory. This process involves multiple data copies, leading to increased latency and reduced performance. The CPU acts as a bottleneck, as it needs to handle data transfer between storage and the GPU. GDS eliminates this bottleneck by directly allowing storage devices to transfer data to and from the GPU’s memory.

We systematically tested every combination of the following parameters in both read and write workloads:

• Block Sizes: 1M, 128K, 16K
• IODepth: 128, 64, 32, 16, 8, 4, 1

As we review our GDSIO results, we examine the read and write performance of the 61.44TB Dapustor J5060 and Solidigm P5336.

GDSIO Sequential Read Performance

The Dapustor J5060 achieves a peak read throughput of 4.2 GiB/s at a 1M block size with IO depths of 64 and 128. At the smallest block size (16K), performance ranges from 0.1 GiB/s to 0.8 GiB/s as IO depth increases. This shows a clear preference for larger block sizes with high IO depths for optimal throughput. The peak performance is achieved at large block sizes, indicating the drive’s efficiency in handling bulk data transfers.

Comparatively, the Solidigm P5336 reached a similar maximum throughput of 4.3 GiB/s at the same block size (1M) but achieved that performance earlier at an IO depth of 32 and maintained it consistently at higher IO depths. This suggests slightly better efficiency in handling large block sizes at a broader range of IO depths for the Solidigm P5336.

To give a better comparative view, we have a differential chart comparing both drives. A greener shade block shows an advantage of the Dapustor SSD, while a block moving to the red side of the spectrum shows a weakness. Here, the J5060 outperforms the P5336 in the 128K block size except for the 4 through 8 IO depths. However, throughput drops are noted at higher IO depths with block sizes 16K and 1M, indicating less efficiency in those scenarios.

In the sequential read latency comparison, the Solidigm P5336 consistently maintains lower latency than the Dapustor J5060 across nearly all block sizes and IO depths. At a 16K block size, the gap becomes more pronounced as the queue depth increases: the J5060 peaks at 2,329 μs at a depth of 128, while the P5336 stays lower at 1,365 μs. At 128K, Solidigm again leads across most depths, with the exception at high loads (4,080 μs on the J5060 versus 5539 μs on the P5336) at depth 128. At the 1M block size, both drives experience latency increases as expected, but the P5336 remains slightly better controlled, with 29,138 μs versus 29,512 μs at the highest queue depth.

GDSIO Sequential Write Performance

The Dapustor J5060 shows a consistent write throughput of 2.7 to 2.8 GiB/s for 128K and 1M block sizes across all IO depths (except 128K, 1 IO depth size, which posted 2.2GiB/s. For 16K block sizes, performance ranges from 0.5 GiB/s to 1.4 GiB/s, depending on IO depth, peaking at 1.4 GiB/s at higher IO depths.

In comparison, the Solidigm P5336 performs better during 128K and 1M block sizes, peaking at 3.2GiB/s. For smaller block sizes (16K), the Solidigm P5336 also shows higher performance, reaching a peak of 1.4 GiB/s at IO depths of 16 to 64. This indicates that the Solidigm P5336 is slightly more efficient with smaller block sizes during write operations.

Moving to a differential view, we see a larger gap opening between the Dapustor J5060 and the write performance of the Solidigm P5336. Our throughput comparison shows that the J5060 falls behind the P5336 in most areas, particularly with large block sizes (1M) across all IO depths. Throughput drops reach -0.5 GiB/s at the 4 IO depths. While there are performance gains at higher IO depths with the 128K block sizes, they are not significant enough to offset the broader underperformance.

When comparing sequential write latency between the Dapustor J5060 and Solidigm P5336, both drives exhibit similar behavior at smaller block sizes like 16K, with Solidigm holding a slight edge at lower IO depths, while Dapustor closes the gap at higher depths (64 and 128). At 128K block sizes, Solidigm again leads at shallow queue depths, but Dapustor consistently delivers lower latency as IO depth increases, indicating better scaling under load. However, with 1M block sizes, Solidigm maintains a clear latency advantage across all IO depths, showing significantly faster response times under heavy sequential write workloads. Overall, Solidigm performs more consistently, while Dapustor’s strength is more visible at mid-sized blocks and deeper queues.

FIO Workload Summary

Flexible I/O Tester (FIO) is an industry-standard benchmarking tool used to measure the performance of storage devices under a wide variety of workload scenarios. Trusted for its versatility and reliability, FIO simulates real-world conditions, providing insights into an SSD’s capabilities and performance limits. StorageReview leverages FIO to offer comprehensive analyses, measuring throughput, latency, and IOPS across workload patterns, block sizes, and queue depths.

Workloads applied:

• 128K Sequential Read and Write
• 64K Random Reads and Writes
• 16K Random Reads and Writes
• 4K Random Reads and Writes

These workloads represent a broad spectrum of enterprise use cases, including large sequential transfers, intensive random I/O typical of databases, and small-block random accesses commonly seen in virtualized environments.

This performance section summarizes the Dapustor J5060’s performance across key synthetic workloads, including sequential and random read/write operations at varying block sizes and queue depths. Metrics are extracted directly from parsed fio output and include bandwidth (MB/s), IOPS, and latency percentiles up to 99.9999%, offering insight into both throughput and tail behavior under load.

128K Sequential Read and Write Performance

The Dapustor J5060 delivers impressive sequential read performance at 128K, reaching 7.48GB/s with tight latency control, even at higher percentiles. Compared to the Solidigm P5336, the J5060’s throughput is essentially the same (7.48GB/s vs. 7.47GB/s). However, Solidigm maintains a slight edge in latency consistency, showcasing marginally lower tail latency.

At 128K sequential writes (QD16), the J5060 achieves solid performance of 3,023MB/s with very low latency. Yet, the Solidigm P5336 surpasses this by a moderate margin, reaching 3,364MB/s, although at a notably higher latency, especially at the 99.99% percentile (4.42ms vs. Dapustor’s remarkably low 0.70ms). This indicates the J5060 is a stronger candidate for latency-sensitive sequential write scenarios.

64K Random Read and Write Performance

In 64K random reads (QD256), the Dapustor J5060 excels with throughput near 7.4GB/s and well-controlled latency. Solidigm’s results closely match (7.47GB/s), with slightly better maximum percentile latency. Both drives perform exceptionally here, with minimal practical differences.

Write performance at 64K random is where the J5060 noticeably struggles, with throughput dropping sharply to 534MB/s and latency rising significantly (742.39ms at 99.99%). In comparison, the Solidigm P5336 significantly outperforms the J5060, delivering 857MB/s and drastically lower latency (221.24ms at the same percentile), making it far better suited for applications sensitive to latency and sustained write throughput.

16K Random Read and Write Performance

At the 16K random read workload (QD256), the Dapustor achieves excellent results with 453K IOPS and controlled latency. The Solidigm P5336 essentially mirrors this performance, slightly edging out the Dapustor in latency (5.47ms vs. 8.16ms at 99.99%), suggesting slightly better latency consistency for Solidigm in heavy random read scenarios.

The Dapustor SSD’s 16K random write performance drops significantly to 32K IOPS, and latency increases to 181.4ms (99.99%). Here again, Solidigm significantly outpaces the Dapustor drive, delivering 51.7K IOPS and a dramatically improved latency profile (71.8ms at 99.99%), underscoring Solidigm’s advantage for latency-sensitive random write workloads.

4K Random Read and Write Performance