performance-comparison-of-containerized-machine-learning-applications-running-natively-with-nvidia-vgpus-vs-in-a-vm

This article is by Hari Sivaraman, Uday Kurkure, and Lan Vu from the Performance Engineering team at VMware. Please find here the original article and all related comments and blog posts.

Performance Comparison of Containerized Machine Learning Applications

Docker containers [6] are rapidly becoming a popular environment in which to run different applications, including those in machine learning [1, 2, 3]. NVIDIA supports Docker containers with their own Docker engine utility, nvidia-docker [7], which is specialized to run applications that use NVIDIA GPUs.

The nvidia-docker container for machine learning includes the application and the machine learning framework (for example, TensorFlow [5]) but, importantly, it does not include the GPU driver or the CUDA toolkit.

Docker containers are hardware agnostic so, when an application uses specialized hardware like an NVIDIA GPU that needs kernel modules and user-level libraries, the container cannot include the required drivers. They live outside the container.

One workaround here is to install the driver inside the container and map its devices upon launch. This workaround is not portable since the versions inside the container need to match those in the native operating system.

The nvidia-docker engine utility provides an alternate mechanism that mounts the user-mode components at launch, but this requires you to install the driver and CUDA in the native operating system before launch. Both approaches have drawbacks, but the latter is clearly preferable.

In this episode of our series of blogs [8, 9, 10] on machine learning in vSphere using GPUs, we present a comparison of the performance of MNIST [4] running in a container on CentOS executing natively with MNIST running in a container inside a CentOS VM on vSphere. Based on our experiments, we demonstrate that running containers in a virtualized environment, like a CentOS VM on vSphere, suffers no performance penlty, while benefiting from the tremenduous management capabilities offered by the VMware vSphere platform.

Experiment Configuration and Methodology

We used MNIST [4] to compare the performance of containers running natively with containers running inside a VM. The configuration of the VM and the vSphere server we used for the “virtualized container” is shown in Table 1. The configuration of the physical machine used to run the container natively is shown in Table 2.

Nvidia

Photo courtesy of Performance Engineering team at VMware

Table 1. Configuration of VM used to run the nvidia-docker container

Nvidia

Photo courtesy of Performance Engineering team at VMware

Table 2. Configuration of physical machine used to run the nvidia-docker container

The server configuration we used is shown in Table 3 below. In our experiments, we used the NVIDIA M60 GPU in vGPU mode only. We did not use the Direct I/O mode. In the scenario in which we ran the container inside the VM, we first installed the NVIDIA vGPU drivers in vSphere and inside the VM, then we installed CUDA (driver 8.0 with runtime version 7.5), followed by Docker and nvidia-docker [7]. In the case where we ran the container natively, we installed the NVIDIA driver in CentOS running natively, followed by CUDA (driver 8.0 with runtime version 7.5),  Docker and finally, nvidia-docker [7]. In both scenarios we ran MNIST and we measured the run time for training using a wall clock.

Nvidia

Photo courtesy of Performance Engineering team at VMware

Figure 1. Testbed configuration for comparison of the performance of containers running natively vs. running in a VM

Nvidia

Photo courtesy of Performance Engineering team at VMware

Table 3. Server configuration

Results

The measured wall-clock run times for MNIST are shown in Table 4 for the two scenarios we tested:

  1. Running in an nvidia-docker container in CentOS running natively.
  2. Running in an nvidia-docker container inside a CentOS VM on vSphere.

From the data, we can clearly see that there is no measurable performance penalty for running a container inside a VM as compred to running it natively.

Nvidia

Photo courtesy of Performance Engineering team at VMware

Table 4. Comparison of the run-time for MNIST running in a container on native CentOS vs. in a container in virtualized CentOS

Please find here the original article from VMware’s Performance Engineering Team and all related comments and articles.

This article is by Hari Sivaraman, Uday Kurkure, and Lan Vu from the Performance Engineering team at VMware

CNIL
Metrics and Logs

(formerly, Opvizor Performance Analyzer)

VMware vSphere & Cloud
PERFORMANCE MONITORING, LOG ANALYSIS, LICENSE COMPLIANCE!

Monitor and Analyze Performance and Log files:
Performance monitoring for your systems and applications with log analysis (tamperproof using immudb) and license compliance (RedHat, Oracle, SAP and more) in one virtual appliance!

Subscribe to Our Newsletter

Get the latest product updates, company news, and special offers delivered right to your inbox.

Subscribe to our newsletter

Use Case - Tamper-resistant Clinical Trials

Goal:

Blockchain PoCs were unsuccessful due to complexity and lack of developers.

Still the goal of data immutability as well as client verification is a crucial. Furthermore, the system needs to be easy to use and operate (allowing backup, maintenance windows aso.).

Implementation:

immudb is running in different datacenters across the globe. All clinical trial information is stored in immudb either as transactions or the pdf documents as a whole.

Having that single source of truth with versioned, timestamped, and cryptographically verifiable records, enables a whole new way of transparency and trust.

Use Case - Finance

Goal:

Store the source data, the decision and the rule base for financial support from governments timestamped, verifiable.

A very important functionality is the ability to compare the historic decision (based on the past rulebase) with the rulebase at a different date. Fully cryptographic verifiable Time Travel queries are required to be able to achieve that comparison.

Implementation:

While the source data, rulebase and the documented decision are stored in verifiable Blobs in immudb, the transaction is stored using the relational layer of immudb.

That allows the use of immudb’s time travel capabilities to retrieve verified historic data and recalculate with the most recent rulebase.

Use Case - eCommerce and NFT marketplace

Goal:

No matter if it’s an eCommerce platform or NFT marketplace, the goals are similar:

  • High amount of transactions (potentially millions a second)
  • Ability to read and write multiple records within one transaction
  • prevent overwrite or updates on transactions
  • comply with regulations (PCI, GDPR, …)


Implementation:

immudb is typically scaled out using Hyperscaler (i. e. AWS, Google Cloud, Microsoft Azure) distributed across the Globe. Auditors are also distributed to track the verification proof over time. Additionally, the shop or marketplace applications store immudb cryptographic state information. That high level of integrity and tamper-evidence while maintaining a very high transaction speed is key for companies to chose immudb.

Use Case - IoT Sensor Data

Goal:

IoT sensor data received by devices collecting environment data needs to be stored locally in a cryptographically verifiable manner until the data is transferred to a central datacenter. The data integrity needs to be verifiable at any given point in time and while in transit.

Implementation:

immudb runs embedded on the IoT device itself and is consistently audited by external probes. The data transfer to audit is minimal and works even with minimum bandwidth and unreliable connections.

Whenever the IoT devices are connected to a high bandwidth, the data transfer happens to a data center (large immudb deployment) and the source and destination date integrity is fully verified.

Use Case - DevOps Evidence

Goal:

CI/CD and application build logs need to be stored auditable and tamper-evident.
A very high Performance is required as the system should not slow down any build process.
Scalability is key as billions of artifacts are expected within the next years.
Next to a possibility of integrity validation, data needs to be retrievable by pipeline job id or digital asset checksum.

Implementation:

As part of the CI/CD audit functionality, data is stored within immudb using the Key/Value functionality. Key is either the CI/CD job id (i. e. Jenkins or GitLab) or the checksum of the resulting build or container image.

White Paper — Registration

We will also send you the research paper
via email.

CodeNotary — Webinar

White Paper — Registration

Please let us know where we can send the whitepaper on CodeNotary Trusted Software Supply Chain. 

Become a partner

Start Your Trial

Please enter contact information to receive an email with the virtual appliance download instructions.

Start Free Trial

Please enter contact information to receive an email with the free trial details.