Unlock Price-Performance and Efficiency on Google Cloud: Databricks & Axion in Action

Summary

• Arm-Based Computing for AI: Maximize performance on Google Cloud using Arm-based Axion C4A virtual machines that deliver breakthrough price-performance and efficiency for the Databricks Data and AI platform.
• Photon Engine Acceleration: Leverage Google's Axion processors to supercharge the Databricks Photon engine, improving query performance and cost efficiency on Google Cloud infrastructure.
• Deployment Optimization: Gain actionable strategies to optimize your Databricks deployments on Google Cloud using the latest innovations in Arm-based computing.

Unlock Price-Performance and Efficiency on Google Cloud: Databricks & Axion in Action

Talk by Mo Farhat, Group Product Manager, Google Compute, Google Cloud (Sponsor Speaker)
Maximize the performance of your Databricks Platform with innovations on Google Cloud. Discover how Google's Arm-based Axion C4A virtual machines (VMs) deliver breakthrough price-performance and efficiency for Databricks, supercharging Databricks Photon engine. Gain actionable strategies to optimize your Databricks deployments on Google Cloud.
Here's more to explore:
Databricks named a leader in the 2024 Gartner® Magic Quadrant™for Cloud DBMS: https://www.databricks.com/resources/analyst-paper/databricks-named-leader-by-gartner
An open, unified approach to your data, BI and AI workloads: https://www.databricks.com/product/databricks-sql
See all the product announcements from Data + AI Summit: https://www.databricks.com/events/dataaisummit-2025-announcements

FAQs

What are Axion C4A virtual machines?

According to this video, Google's Arm-based Axion C4A virtual machines deliver breakthrough price-performance and efficiency for Databricks workloads on Google Cloud.

How does Axion improve Photon engine performance?

This video demonstrates how Arm-based Axion processors supercharge the Databricks Photon engine, improving query performance and enabling better price-performance for analytical workloads.

What are the cost efficiency benefits?

By using Arm-based Axion C4A VMs with the Databricks Data and AI platform, organizations can achieve significant cost savings while maintaining or improving query performance.

How do I optimize my Databricks deployment on Google Cloud?

This video provides actionable strategies for optimizing Databricks deployments on Google Cloud, including VM selection, Photon configuration, and workload placement best practices.

Full transcript

Welcome everyone. I'm a product manager in Google Compute Engine and this is the data bricks uh Axion and action session. So we're going to be talking about Axion and uh pleased to be joined here by Gri Shakorti from epsilon and Abashakry senior director of engineering at data bricks and uh we'd love to uh have your attention today. Any questions? U we'd love to answer those as well. All right, so let's get started. U first uh the introduction page.
Um so first let me start off with some feedback we hear from customers. Um, customers tell us that AI is becoming an essential part of their businesses today. And then AI is leading to more immersive applications that have higher requirements, performance, and more specialized virtual machine needs. and customers second that are well on their way in their modernization journey tell us that they're building now for even more scale even more efficiency and looking to use more of the emerging cloud services uh that are on offer today.
And then finally of course for traditional enterprises traditional enterprise workloads homegrown applications remain bedrock core of many businesses today. we cannot lose sight of the high uh reliability and stability requirements that those applications need. Right? And uh like spend a little bit of time about how we build our approach here at Google Comput Engine. We analyze workloads across compute, storage, networking, across the hardware and software stack.
We work with customers to understand deeply their needs, their problems, the bottlenecks that they're seeing with workloads, and then we design workload optimized infrastructure to solve these problems. And so, just want to spend a a couple minutes here talking about, you know, the um underlying technology that powers all this innovation at Google. Google has been a pioneer in data center designs, pioneered many data center technologies that power the planet scale services that customers rely on today.
Search of course, maps, YouTube, Gmail, others. And at Google Compute Engine, we make this innovation available for customers so that you too can run like Google. And underlying all of this is the Titanium system of technologies. Titanium is a custombuilt networking security and tiered offload system that scales performance across bottlenecks in compute, storage and networking. And on the compute side, we've um continued to deliver leading price performance virtual machines.
Well, latest processors from Intel, AMD and of course now Google Axion. On the storage side, uh our next generation uh offering is called Hyperdisk, which helps you scale storage performance independently of the VM up to 500K IOPS. a huge number with hyperdisk extreme volumes. And on a networking side, of course, we continue to innovate there with the Titanium ML adapter based on CX7 that accelerates machine learning workloads with our cloud RDMA technologies that are in our latest high performance computing instances up to 200 Gbits per second and ultra- low one-way latencies less than three microsconds uh to help scale high performance computing clusters there.
And um with that uh we now have been investing in expanding the general purpose uh portfolio and the compute VMs portfolio at Google um to meet emerging needs. This expansion continues. Uh we're at a place now where we offer a complete um portfolio of machines there to serve every workload. Um starting with the flexibility of the N and E series VM families, consistent high performance, advanced maintenance controls of the C family of VMs and this uh expansion also continues for specialized VMs and now offering of course high performance computing uh options with the latest being H4D, storage optimized offerings with the Z family of instances, memory optimized with M and X family of instances. and GPU accelerated machines with the GNA family instances including the latest uh B200 and GB200 offerings that are now available on Google Cloud.
And so let's let's dive in a bit into the C family and C4A and Axion, right? So um first just a little bit of context here. Why did Google build on build built Axion on ARM and why should you build on Axion? So, Axion today provides the best price performance for the growing ecosystem of ARM compatible workloads and provides the best performance per vCPU for key applications there in database analytics common serving workloads and it does this with incredibly efficient performance per watt to help you reduce the carbon footprint of your operations.
And so, this here is Google Axion. Uh we launched it at uh cloud next last year and it is Google's first custom server processor. It is the world's fastest ARMbased processor today with up to 10% better performance than the leading current generation ARMbased instances you can find in the cloud and up to 65% better price performance. up to 60% better energy efficiency than the current generation x86based instances on Google cloud today and just key here that we built this machine on standard ARM compute cores and we did this to give you access low friction access to the expanding ARM compatible software ecosystem but also to make it incredibly easy for you to migrate your workloads from existing ARMbased instances in the cloud to Axion and run ARM multicloud as well and to do so with minimal to no effort right so um to just dig a little bit deeper into the VMs uh themselves here so the Axion's available uh powers the C4A virtual machine family C4A virtual machines provide complete range of general purpose sizes to 72 vCPUs 248 gigabytes per of memory uh per vCPU to 6 TB of uh local SSD storage again based on titanium SSDs um that provide 35% lower latencies um particularly beneficial for database workloads and analytics workloads and um up to 100 gigabits per second and tier one networking um of course all built on titanium and full support for hyperdisk uh balanced and extreme today throughputs coming very So all right so I'll end with this.
Please try C4A. Try C4A on Axion. I think you'll be very impressed. And please welcome. Welcome. Thank you Mo. Hello everyone. Uh. I'm a VP at Epsilon. Epsilon is part of publicist group who is world's largest marketing providers at this point of time. What epsilon does is people ask epsilon a simple question in terms of marketing. How can I improve ROI, return on investment of marketing dollars? To do that, epsilon performs certain operations.
One is identify the real person whom do they need to reach. Second thing is at what time they should reach and what message they should give. And third is what are the messages what are the consistent messages they need to provide customers so that you can convert them as the that particular brand's customer. So that's the overall marketing spin and how we do that I'll just go through in a glance. So on the left hand side you will see different kind of a data source when we do the marketing we actually market to real customers and that's where epsilon's power of identity comes into picture where we can identify a real person to send the right message at a right time so that brand can acquire themsel in the middle you see different services of epsilon uh like CDP which is a customer data platform it has all the customer data else information.
We have clean room where along with datab bricks we can find out we can activate the customers we can build an audience in a privacy safe manner. Everything epsilon does is in a privacy safe manner and every brand we can connect their first party data to epsilon's own data set to provide a much larger reach. Then we use our messaging solutions to send the message. It can be either email message, personalized email, text or display messages in your connected TV or anywhere using multi- channelannel messaging capability and then we can use activation through own channel or paid channel it can be epsilon digital or uh it can be even YouTube social media any other channels and at the end on the right side and in the below We are showing the entire databicks lakehouse platform which powers the solutions.
We use datab bricks data lake. We make sure all of the objects are optimized using liquid clustering, predictive optimizations enabled and databicks lakehouse provides a complete entire picture using auditing data lineage everything in one get. Now when we talk about all these things, gravity like ARMbased architecture plays a significant role in epsilon's data processing and we process the data in multiple clouds um in all three hyperscalers basically and that's where I will talk about the use case how ARMBS architecture helps to solve this solutions in this particular use case where we use a uh it's kind of a data warehouse use case.
The P was over a 65 terabyte of data set. It has a significant amount of joins and large aggregates built in. The result there it has large amount of shuffle both shuffle read and shuffle write operations. We will go through why it is important and how basically this new aion systems are helping us. And this has been run by several data scientists for ML jobs because we are using one lakehouse architecture where the data lake where all the ML job runs and data warehouse where all the BI analytics job runs.
It runs on a same data source so that we have one system of truth and we are all these use cases we are comparing with both x8664 architecture along with newly released ARMS ARM 64 architecture machines in GCP. From a runtime perspective, uh data bricks we use 154 runtime in this use case. U but eventually we will also use 164 DBR mlr as applicable and AQE which is part of the again datab bricks features which helps significantly basically wherever we have a data skew AQE helps a lot in this kind of a scenario.
As I mentioned all tables are already using liquid clustering. So here is the result. Couple of things I wanted to highlight here. Uh number one we used x8664 in GCP versus ARM 64. Two things I wanted to segregate very specifically here is one we used ARM without local SSD and the second one we use ARM with local SSD. There is a significant difference in performance with or without local SSC. I'll explain it why. The data set which epsilon executes in most of our marketing applications we are talking about multipetabyte data set and over here there is a significant amount of shuffle happens there are spill happens autoscaling of the disk devices happens when we use local SSD on that the overall performance and throughput what you will get is substantially higher and your runtime drastically reduces.
This is one of the area that's why you will see the first use cases with local SSD and the second use case without local SSD there is a significant difference but the important thing to note here is in the x8664 architecture when we used even with local SSD ARM architecture with a C4 instances it's giving us still more than 20% performance improvement and anywhere between 10 to 15% cost advantage. So that's the key and we have seen this C4A instances are providing possibly the best result across all of the cloud platform that we have run.
So this is overall observations of our use cases is like we tested it two months back when uh it was released and data bricks search started supporting it and overall performance in terms of the throughput and in terms of what we need is really helping when we are using C4A architecture. Now one thing to note as well is depending on the application you run it needs to also be compatible with ARM. Not every application is compatible with ARM.
So that's that's one use case which Abishek may u touch up later on. So this is uh the overall impact because of the using local SSD on top of ARM. uh we are getting significant speed up in terms of uh shuffle read and write and that gives a massive boost in our application throughput. This is the final story. the epsilon's unified lakehouse solution uh what we are deploying in multiple cloud and also in GCP where this entire C4A instance plays a key role over here you will see all of in the center is epsilon's common data set uh which includes identity which you you includes uh all of the various attribute informations which we connect with customers firstparty data in a privacy safe manner in customer's own data own area.
Think of it in a GCS storage for customer and all of the compute is also dedicated for that customer. So there is no data sharing or anything cross customer and this powers basically all of the application being it clean room being it CDP loyalty or any activation platform retail media network messaging you can think of it's a collection of lakehouse connected lakehouse which can talk together to provide all of the features that a customer needs to make sure the campaign is successful they can start the campaign they can build audience.
They can activate audience. They can measure real time and figure out if the campaign is performing well or not. If not, they can change it dynamically without waiting for the entire campaign to end and giving them not the satisfactory result. So they have all these things and all of these options are flexible for customer to choose at any point of time in any cloud of their choice and this is what we are deploying at GCP.
Now I will hand over to he will take us through to the rest of the parts using databicks platform. Cool. Thanks. Hey everyone this is Abhishek Ray from data bricks. uh I'll be continuing the story on Exxion and share our perspective on what we are seeing with Exxon across our customers. First a quick recap of data bricks itself. Uh as you might be aware data bricks primarily serves three use cases across our user community.
Data analytics for data analysts, data engineers, data scientists and IML engineers. data brick SQL warehouse which is a more than $500 million business today. It's our fastest growing product and Genai where we are rapidly growing our user base. It is our point of view that Genai will not succeed without governance and that's where Unity catalog comes in probably heard plenty about that in DAS this time also quick recap of Photon which is essential to our Exxon story.
So, Photon is the fastest native uh performance uh engine built from the ground up for modern cloud hardware. It provides up to 80% better performance and up to 12x better TCO. Most importantly, Photon is 100% compatible with Apache Spark. So in order to benefit from Photon, no code changes are required. And finally, Photon is built to support all use cases, all data use cases. So it is really a drop-in replacement for Spark.
With that context, let's move on to key factors for us to select Exxion. First of all, Exion offers pretty significant performance improvements right out of the box. up to 40% better compared to second gen instances. The ARMbased Exxon cores they are designed from the ground up to provide maximum benefit for data inensive workloads the similar to what data bricks runs. We are also very excited about the improvements in the storage stack that come with the Exxon platform.
Specifically, hyperdisk storage allows you to independently scale performance and size, which is a great win for TCO and performance. And we also have observed that as I'll cover in the next slide the combination of titanium local SSDs as well as u hyperdisk storage like what Garrick mentioned that is really very helpful for data intensive workloads like data bricks ultimately helping with both performance and TCO. We're also very excited about the networking stack in Exxion because it's able to support encryption at full line rate. uh something that's very important for us from enterprise security point of view.
And on that note, Exxon overall fits in very nicely with our data security architecture. With that, let's go over the results of uh our our experience running on Exxion and having helping our customers adopt Exxion over past several months. First of all, Exion really loves sorry lives up to its promise. We are seeing uh performance gains up to about 50% compared to second gen instances and about 40% compared to third gen instances.
These performance gains most importantly come without any code changes or deeper optimizations that anyone has to make. Another important call out is that in addition to the performance gains, there are very consistent TCO gains as well that we are seeing across the board and that also uh is a pattern we are seeing across a whole range of workloads. A whole range of workloads from computer intensive, data intensive, everything gains with the Exxion platform in a very consistent fashion including IObound workloads.
So the examples that Garrick mentioned earlier where you see a difference between local SSD and hyperdisk uh performance the huge jump there a lot of that signals like IObound workloads uh with hyperdisk storage we are seeing that uh hyperdisk and titanium local SSD combination we are we are seeing pretty consistent improvements data bricks in has a capability called uh uh disk autoscaling where we are able to dynamically grow your remote block storage to match your storage and IOPS needs and that is very handy in ensuring that you're not overpaying for IOPS.
You're mainly paying proportional to the amount of compute you are using. Uh in fact uh one of the nice things with the hyperdisk uh stack is that uh we've found the flexibility so powerful that we are actually backporting that support to third gen instances where we previously did not support hypertext storage. Also one important call out is that u it's ultimately the hardware and software combination that you have to take together to maximize your wins. uh as the chart on the screen shows the maximum gains come with Exxon on data bricks when using the data bricks photon engine the native execution engine I mentioned earlier u so as a concrete example from this you can see that uh when we move from uh data bricks photon engine on n2 to photon engine on c4a we get about 50% performance improvement Going from Spark on N2 to Photon on C4A is a 10x performance improvement and you know in hindsight that's not really a surprise because we really love the Exxon platform the Titanium platform architecture that Mo described earlier and data bricks Photon engine also is really optimized for native native execution on such a modern platform.
One important call out is that uh when used with data bricks uh the Java version does matter. Java received many important ARM related upgrades between earlier versions and JDK 17 and onward. So that's something you want to take along. Just another quick deep dive on some of the benchmarks. So this slide shows our end toend benchmarks as well as some microbenchmarks. Specifically, it compares Exxion or C4A performance compared to third generation x86.
You can see pretty broad improvement across the board. Most importantly, like I mentioned earlier, no code changes, no optimizations. It's exactly the same code running on both platforms. In fact, uh it was this result that really simplified our choices with respect to a recommending Exxion to our customers and b ourselves adopting Exxion and data bricks own serverless offering. So we are going full throttle with Exxion in our serverless offering on GCP. data brick serverless just to recap is uh uh a platform that's designed to provide much better customer experience and TCO through cutting edge uh intelligence built into the platform and several other optimizations that we've built in every single product of data bricks integrating into serverless so that is going full exon finally I'll leave you with a few tips uh first of all uh you know exon is great platform as the capacity is coming up build for both x86 and ARM and I think that might be true even in the long run that'll just give you maximum optionality for data bricks customers adopt fleet fleet is our flexible offering for adopting for uh falling back to whatever instance types are available I'll cover that in the next slide in a little more detail don't ignore the software like I mentioned earlier so Java was one example where use a newer JDK version that will help you tremendously and for data bricks customers that would mean using a newer version of data bricks runtime with photon engine.
Finally, embrace hyperdisk. It is really transformative. The ability to separately tune performance and size is uh super helpful. Uh with that just a quick note on the fleet offering I mentioned earlier. So data bricks fleet is in private preview on GCP. It provides two broad capabilities. Number one, it frees customers from having to worry about uh capacity issues or uh configuring uh fallbacks at your own job configuration level when the preferred skew is not available.
Now you can push your capacity requirements into datab bricks in terms of the fallback skew order for example. In addition, datab bricks can automatically pick the right instance queue for you. And that's something that our users use very heavily. The users who onboarded and the users who've used it before on AWS where we've had the fleet offering before. Another very important win of the fleet offering which would strongly recommend uh using that as well is it significantly improves spot usage on data bricks.
So normally on data bricks you when you're using spot in most cases you have spot with fallback to on demand that's configured but with fleet you can effectively have a dozen different options of spot tried before you give up on spot eventually for example. So that way adopting fleet will give you the maximum optionality being able to leverage ARM and when not available you can fall back and also it'll help you maximize spot availability across whole range of SKs.
That's all I had. Thank you.

Learn more about the Databricks Data and AI platform.