8 MINS READ
What we do
Highlights
- Virtual machine (VM)-based platforms perform better for static and consistent workloads, whereas serverless platforms maintain a steady response time, irrespective of the increase in workloads.
- VM platforms are better for long-duration workloads with higher resource utilization, whereas serverless platforms are ideal for short-duration workloads with low resource utilization.
In this article
In this articleinpage
Cloud migration
Cloud services and software-as-a-service help build high-performance digital enterprises. With the increasing adoption of artificial intelligence, organizations need to consider how to deploy machine learning (ML) and deep learning (DL) workloads. Cloud migration offers two options—virtual machines (VM) or serverless cloud deployment. Here are a few points to consider for high performance at optimal costs:
Virtual machine-based platforms perform better for static and consistent workloads
Serverless platforms maintain a steady response time, irrespective of increase in workloads
VM platforms perform better for long-duration workloads with higher resource utilization
Serverless platforms are ideal for short-duration workloads with low resource utilization
ML platforms are easy to deploy but harder to scale as operations grow
Serverless platforms are more scalable and better for ‘bursty’ or variable workloads
Serverless computing is more cost-effective and may be the best way forward to support complex AI applications.
The future of digital enterprises will be driven by artificial intelligence (AI). As companies embrace AI, they are also investing more in cloud services and software as a service which make it easier to manage machine learning (ML) and deep learning (DL) workloads. However, deploying ML and DL workloads on the cloud incurs a recurring cost. In addition, cloud deployment comes with service level agreements (SLAs) to deliver low latency and high throughput, especially while inferencing the models for making predictions.
To ensure optimal workload deployment, organizations must choose cloud services that ensure high performance and are cost-effective. One approach is to use containers with virtual machines (VMs) or optimally configured instances in ML platforms. Serverless-architecture-based deployment is also a popular mechanism for cloud migration due to its high scalability and pay-per-use cost model. This white paper highlights the cost and performance tradeoffs between an ML platform and serverless deployment. This knowledge will be highly beneficial in deploying applications such as recommender systems.
Workload deployment on cloud: Balancing cost and performance
On-premise deployment of applications requires an upfront investment to build compute capacity. In certain instances, due to peak workload, purchasing, and deploying hardware resources in excess to avoid SLA violations, results in higher maintenance costs. This is more likely to happen when resources are idle due to low load during non-peak time.
To avoid this, enterprises are migrating their ML and DL workloads to the cloud since it offers a highly flexible infrastructure, low operational management, and better performance. Cloud vendors also provide many services, platforms, and hardware instances for deploying these workloads. For enterprises, the choice should be based on application characteristics and design, performance requirements, and the available budget. This will ensure the right balance between cost and performance.
Many popular cloud vendors provide ML platforms for deploying AI workloads. These platforms ease the process by providing automatic training,deployment, and inference of the models.
These platforms from cloud vendors also provide various computing, memory-intensive CPU (central processing unit), and GPU (graphics processing unit) instances for deploying ML and DL models. Defining appropriate scaling policies helps address SLA violations due to dynamically varying workloads. However, frequent scaling up and down the VM instances due to ‘bursty’ or variable workload could increase the cost and degrade application performance.
Serverless platforms are emerging as an alternate option for applications built using models based on ML and DL for inference. The autoscaling feature in the serverless platform allows up-scaling and down-scaling instances as the workload increases or decreases. Moreover, billing is based on the actual usage, unlike VM billing instances, where the cost depends on the lease duration, irrespective of the actual usage. However, there are certain inherent constraints in serverless platforms, such as statelessness and limited compute power. In addition, serverless instances have a 15-minute lifetime and suffer from a cold start problem. This results in high latency for the first inference request served by the model. Some of the popular cloud vendors and open-source platforms offer serverless platform services.
THE RIGHT CLOUD DEPLOYMENT PLATFORM FOR AI APPLICATIONS: A USE CASE
The following use case of a recommender system will help understand the tradeoffs between cost and performance for an on-premise, ML platform, and serverless architecture-based deployment of an AI application. A recommender system based on a graph neural network uses products chosen in the past to recommend the next product. The three scenarios for analyzing deployment of the recommender system are as follows:
On-premise servers containing four and eight physical cores with 256 GB memory.
ML platform with ml.c5.xlarge (4-cores) and ml.c5.2xlarge (8-cores) compute family CPU instances and gdn.ml.xlarge and gdn.ml.2xlarge family GPU instances.
Serverless instances with 3GB memory resulting in two cores on each instance.
At a lower workload, the on-premise deployment results in the least response time compared to the ML and serverless platform. However,as the workload increases, the response time increases for on-premise and ML platforms. This is primarily due to the fixed number of resources and compute capacity (number of cores) available for on-premise servers and slow scalability of the ML platform. This is primarily due to the fixed number of resources and compute capacity (number of cores) available for on-premise servers and slow scalability of the ML platform.
The response time for serverless architecture remains the same with an increase in workload.
On the other hand, response time for the serverless platform remains steady even with an increase in workload. This is because, with an increase in inference requests, the number of instances serving these requests increases proportionally. Moreover, new serverless instances are spawned faster compared to those on the ML platform.
One of the challenges of cloud migration is comparing the cost efficiency of different services across varying cost models. The cost of ML platform instances is fixed and independent of the load served. In the serverless option, costs change with the workload and new instances. Let's consider the cost comparison of instances used for inference in the ML and serverless platforms in the graphic below. The ML platform is cost-effective when the request rate served by instances is high. However, a serverless platform is a better choice if the resources reserved for inferencing are underutilized. The intersection of the serverless platform with the instances represents the breakeven point. ML platform instances can serve workloads with request rates above this breakeven point with low latency and cost. Conversely, workloads with a request rate below the intersection point means that the ML platform instances are underutilized, resulting in higher costs. In such a scenario, the serverless platform is a cost-effective option due to its pay-per-use cost model.
From a cost perspective, a serverless platform is a better choice if the resources reserved for inferencing are underutilized.
Assessing the performance-cost ratio across multiple dimensions
There are certain scenarios where an on-premise and ML-platform-based deployment delivers better performance and cost ratio. For example, when a workload consists of many requests and resource utilization remains high for a long duration. Serverless platforms work better for short-duration workloads with low resource utilization.
The on-premise and ML-platform-based deployment on cloud platforms performs better than serverless platforms for smaller workloads. As the workload increases, the application performance drops due to limited scalability and resource constraints. Such workloads can be migrated to the cloud to leverage the extensive and flexible infrastructure. ML platforms on the cloud are easy to deploy but are slow when it comes to scaling operations. Serverless platforms provide high scalability and result in cost-effective deployment, particularly for ‘bursty’ or variable workloads.
The choice of the ideal cloud deployment platform should be based on application characteristics, which extract high performance at optimal costs. Compute requirement, memory usage and expected workload are some parameters for comparison.
The use of serverless platforms in conjunction with ML platforms might be advantageous for dynamically varying workloads. In this approach, the workload can be balanced between ML platform and serverless instances based on the instantaneous utilization of the resources, as depicted in the graphic below.
Load balancing across cloud services
The future of AI lies in serverless computing
When deployed correctly, serverless computing is ideal for handling ML and DL workloads, offering dramatic cost savings due to execution-based pricing and better resource utilization. It enables rapid prototyping of new and complex AI services by eliminating the time spent on managing the server’s infrastructure. In the future, cloud vendors may provide GPUs within serverless platforms which would increase computation power significantly and deliver state-of-the-art AI capabilities. Furthermore, it will support compute-intensive jobs such as training large deep learning models. Going forward, a serverless architecture may be the driving force behind AI innovation.
