SPECstorage™ Solution 2020_ai_image Result
Copyright © 2016-2024 Standard Performance Evaluation Corporation
|
SPECstorage™ Solution 2020_ai_image ResultCopyright © 2016-2024 Standard Performance Evaluation Corporation |
| Qumulo, Inc. | SPECstorage Solution 2020_ai_image = 704 AI_Jobs |
|---|---|
| Azure Native Qumulo - Public Cloud Reference | Overall Response Time = 0.84 msec |
|
|
| Azure Native Qumulo - Public Cloud Reference | |
|---|---|
| Tested by | Qumulo, Inc. | Hardware Available | May 2024 | Software Available | May 2024 | Date Tested | May 11th, 2024 | License Number | 6738 | Licensee Locations | Seattle, WA USA |
None
| Item No | Qty | Type | Vendor | Model/Name | Description |
|---|---|---|---|---|---|
| 1 | 33 | Microsoft Azure Virtual Machine | Microsoft | Standard_E32d_v5 | The Standard_E32d_v5 Azure Virtual Machine runs on the 3rd Generation Intel® Xeon® 8370C processor. Each VM features 32 vCPU cores, 256 GiB of memory, 8 NICs, and supports up to 16000 Mbps of network egress bandwidth. Network ingress is not metered in Azure, enabling exceptional throughput results of 68GB/s with an overall response time latency of 0.84ms. Thirty-two (32) VMs served as Spec netmist load generators, each with two dedicated ANQ mount points. One (1) VM was dedicated to the Prime Client role. Each VM utilized a single NIC with Advanced Networking enabled. |
| Item No | Component | Type | Name and Version | Description |
|---|---|---|---|---|
| 1 | Azure Native Qumulo filesystem | File System | 7.1.0 | The Azure Native Qumulo (v2) File System is a cloud service provisioned directly from the Azure portal. This architecture differs from previous releases of Qumulo Core Filesystems. Customers specify the initial capacity and a customer subnet delegated to Qumulo.Storage/FileSystems. Qumulo uses vNet injection, resulting in IP addresses on customer networks. The storage service components are opaque to the customer. In ANQ, performance and capacity are completely disaggregated. The service offers non-disruptive and dynamic performance increases on-demand, currently requested through service requests. For this benchmark, an increase in aggregate performance capability up to 100GB/s of throughput and 500,000 total IOPS was requested. |
| 2 | Ubuntu | Operating System | 22.04 | The Ubuntu 22.04 operating system was deployed on all Standard_E32d_v5 VMs. |
| Component Name | Parameter Name | Value | Description |
|---|---|---|
| Accelerated Networking | Enabled | Microsoft Azure Accelerated Networking enables single root I/O virtualization (SR-IOV) |
N/A
| Component Name | Parameter Name | Value | Description |
|---|---|---|
| mount options | 64 | Each NFSv3 client mounted (2) distinct IP addresss |
The single filesystem was attached via two (2) NFS version 3 mounts per client. The mount string used: mount -t nfs -o tcp,vers=3 10.0.0.4:/AI_IMAGE /mnt/AI_IMAGE-0 mount -t nfs -o tcp,vers=3 10.0.0.5:/AI_IMAGE /mnt/AI_IMAGE-1 The mount specifies that two (2) NFS version 3 exports were mounted on each client.
Azure Service SLA: https://www.microsoft.com/licensing/docs/view/Service-Level-Agreements-SLA-for-Online-Services
| Item No | Description | Data Protection | Stable Storage | Qty |
|---|---|---|---|---|
| 1 | Azure Native Qumulo persists data entirely using Azure Blob Storage, which is a Microsoft-managed service providing cloud storage that is highly available, secure, durable, scalable, and redundant. Architecturaly, Read cache is serviced from an in-memory L1 cache and a NVMe based L2 cache. The global read-cache is increased on-demand with a customer service request. Write transactions leverage high performance Premium SSD v2 disks, which act as a protected write-back cache for incoming writes, continuously flushing data to Azure Blob Storage. | Provided by Microsoft Azure Blob Storage | Azure Blob Storage | 64 |
| Number of Filesystems | 1 | Total Capacity | 100 TiB | Filesystem Type | Azure Native Qumulo |
|---|
Azure Native Qumulo is deployed directly from the Azure Portal, the Azure CLI, or using an Azure REST endpoint. Customers specify the capacity of storage and what subnet they want to deploy vNet-injected IP addresses in.
Azure Native Qumulo is a cloud storage service with a different architecture than traditional storage, including the traditional Qumulo Core software. Data is persisted to Azure Blob, with caching occurring only at the compute layer. This disaggregates storage capacity/persistence from compute/performance into two completely separate and independently elastic layers. The Azure Native Qumulo filesystem persists data to Azure Blob using Locally Redundant Storage (LRS). Azure Blob LRS stores data synchronously across three physical locations within an Azure region. This architecture acts as an accelerator, executing parallelized reads that are prefetched from object storage and served directly from the filesystem’s L1/L2 cache to the clients (over NFSv3 in this benchmark).
| Item No | Transport Type | Number of Ports Used | Notes |
|---|---|---|---|
| 1 | Ethernet | 97 | There were ninety-seven (97) vNics used in this benchmark, with thirty-two (32) used as the load generatoring Ubuntu NFS clients, sixty-four (64) used as vNet injected IP addresses from the ANQ storage service, and one (1) by the Prime Client. Each VM utilized a single NIC with Advanced Networking enabled. |
| 2 | ethernet | 97 | Azure networking allows up to 8 virtual NICs for the Standard_E32d_v5 with a combined aggregate EGRESS limit of 16,000 Gbps. In this benchmark only a single vNic was used, which still has the same EGRESS limit of 16,000 Gbps. There is not a INGRESS limit in Azure. |
The NFSv3 protocol was used with the mount options tcp and vers=3. There was no use of nconnect in this benchmark, as it does not result in a performance increase with this workload. IP addresses are distributed to clients using a round-robin DNS record, as shown in the architectural diagram. Each client in this benchmark used two of the ANQ vNet-injected IP addresses per NFSv3 client, resulting in two distinct NFSv3 mounts on each load-generating client.
| Item No | Switch Name | Switch Type | Total Port Count | Used Port Count | Notes |
|---|---|---|---|---|---|
| 1 | Azure | 12.5 Gbps Ethernet with Accelerated Networking | 64 | 64 | Used by ANQ Storage Service |
| 2 | Azure | 16.0 Gbps Ethernet with Accelerated Networking | 33 | 33 | Used by NFS Front-end Clients |
| Item No | Qty | Type | Location | Description | Processing Function |
|---|---|---|---|---|---|
| 1 | 1 | vStorage | Azure Region EastAsia | Azure Native Qumulo Cloud Storage Service | Qumulo File System, Network communication, Storage Functions |
None
| Description | Size in GiB | Number of Instances | Nonvolatile | Total GiB |
|---|---|---|---|---|
| Azure Standard_E32d_v5 VM memory | 256 | 33 | V | 8448 | Grand Total Memory Gibibytes | 8448 |
Each Standard_E32d_v5 VM has 256GiB of RAM
Azure Blob Storage provides the stable storage function in the ANQ architecture. It is a cloud object storage solution from Microsoft Azure, designed to store large amounts of unstructured data and optimized for efficient storage and retrieval of large data objects. This solution offers virtually unlimited scalability, durability, and high availability to meet ANQ’s stable storage needs. From a data protection perspective, Azure Blob Storage uses HMAC to ensure data integrity. When data is uploaded, a hash is computed and stored. Upon retrieval, the hash is recomputed and compared to verify data integrity. ANQ employs Azure Blob Storage’s Locally Redundant Storage (LRS) offering, which keeps multiple copies of data within a single region.
The solution under test was a standard Azure Native Qumulo cluster deployed natively in Azure. The cluster is elastic both in performance and capacity and can be requested to handle 100's of GB/s of throughput, millions of IOPS, and 100's of PB's of capacity. Performance and capacity are 100% disaggregated in ANQ.
For this benchmark, the aggregate performance capability of 100GB/s and 500,000 IOPS was requested for the ANQ. Performance capability can be changed up or down by submitting a service ticket, at this time.
None
None
None
Generated on Sat Jun 8 13:32:46 2024 by SpecReport
Copyright © 2016-2024 Standard Performance Evaluation Corporation
