Skip to main content
Announcements
Have questions about Qlik Connect? Join us live on April 10th, at 11 AM ET: SIGN UP NOW
cancel
Showing results for 
Search instead for 
Did you mean: 
PedroC
Contributor II
Contributor II

AMD EPYC dual sockets servers

Hi

I'm planning to upgrade my enterprise qlik sense servers and I've noticed that Qlik now includes AMD EPYC processors in their whitelist but in single socket configurations. 

I was wondering if anybody has tested Qlik Sense in a dual socket AMD EPYC configuration. 

Dual socket configurations has been in the machine whitelist but only with Inter cpus. 

Thanks in advance.

 

1 Solution

Accepted Solutions
Levi_Turner
Employee
Employee

Our Scalability labs haven't currently tested the third generation EPYC chips (of which the 7763 is a part of). That being said, the colleague who does Scalability work had these thoughts:

  • The main architectural differences should benefit Qlik products
  • Very little risk in using this in a 1 socket configuration

Personally? One does need to consider the trade off of Intel vs. AMD in these areas. AMD's calling card with their revamped EPYC series is a crap ton of cores with a lower clock speed. Intel has been struggling of late to keep up with AMD in next generation fabs, but their chips have higher clock speeds.

When it comes to performance in Qlik, super high end performance in unsaturated environments will perform better if the clock speed is higher, right? So if you wanted to ensure sub X second response time for a handful of apps, Intel will likely outperform AMD. In my experience, this is a niche use case. The other area where clock speed is extremely relevant is application reloads.

Most environments have variable compute needs. They add apps year over year. Developers make mistakes and SELECT * from a large database. So on and so forth. In those variable work-load environments, AMDs shine due to the volume of cores that they provide. Not because the per request speed is necessarily faster in isolation, but because the summation of all request speed is fast when there's more cores to handle the work units especially when there is resource contention.

View solution in original post

5 Replies
Levi_Turner
Employee
Employee

This recently came up and was chatting with one of the folks on the team which produces the whitelist / recommended servers. From them (with slight edits for clarity):

I do have a two socket EPYC in our lab. .. What I can say from the preliminary test results is that I wouldn’t recommend it at this moment. .. the internal architecture of the EPYC chips make a dual socket system look similar as an 8 socket system in the Intel world. So the results I got right now are that in very few cases the dual socket system will have better response time than the single socket system with the same CPU. In most cases the single socket with the same CPU will perform better.


In brief, 2 socket EPYC solutions are not only not whitelisted but are expected to perform notably worse than the same chip in a single-socket configuration.

Cheers

PedroC
Contributor II
Contributor II
Author

Thank you Levi

It's just what I expected. Qlik does include Intel dual socket CPUs in their white list, in fact our production environment is supported by a Dell PowerEdger R630 with two Intel Xeon E5-2690 v4 (14 cores at 2.6GHz).

It does perform quite well but the number of concurrent users has been increased along the last 4 year and it's time to upgrade.

Our Dell representative offered us a  Dell Poweredger R740 with two Intel Xeon Gold 6248R (24 cores at 3 GHz) but we have some doubts about it (if we compare cpumark of both cpu we are not getting that much performance for the value)

That's why we started thinking about AMD EPYC.

Right now I'm evaluating a Dell Poweredge R75250 with one AMD EPYC 7763 (64 cores at 2.45 GHz). Checking cpumark is about 40% better than the Intel offered with two socked so I think it's a better option.

Do you have any feedback about the AMD EPYC 7763 with 64 cores?

Thanks in advance

 

 

Levi_Turner
Employee
Employee

Our Scalability labs haven't currently tested the third generation EPYC chips (of which the 7763 is a part of). That being said, the colleague who does Scalability work had these thoughts:

  • The main architectural differences should benefit Qlik products
  • Very little risk in using this in a 1 socket configuration

Personally? One does need to consider the trade off of Intel vs. AMD in these areas. AMD's calling card with their revamped EPYC series is a crap ton of cores with a lower clock speed. Intel has been struggling of late to keep up with AMD in next generation fabs, but their chips have higher clock speeds.

When it comes to performance in Qlik, super high end performance in unsaturated environments will perform better if the clock speed is higher, right? So if you wanted to ensure sub X second response time for a handful of apps, Intel will likely outperform AMD. In my experience, this is a niche use case. The other area where clock speed is extremely relevant is application reloads.

Most environments have variable compute needs. They add apps year over year. Developers make mistakes and SELECT * from a large database. So on and so forth. In those variable work-load environments, AMDs shine due to the volume of cores that they provide. Not because the per request speed is necessarily faster in isolation, but because the summation of all request speed is fast when there's more cores to handle the work units especially when there is resource contention.

PedroC
Contributor II
Contributor II
Author

Thanks again Levi

I agree with you almost every point of your reply. Right now we do have some apps with SELECT * and huge amount of data. Those apps are QlikView and we are migrating them to Qlik Sense with a different approach. Instead of few huge apps (5-6 GBytes each) with dozens of tables we a creating dozens of apps way below 1 GBytes using Sense.

Of course, performance is way better.

 

The point is that we still have two huge QlikView apps used by more than 300 users (usually no more than 50 concurrent users) and users are suffering poor performance and eventually severe lags.   

That's why we're planning to move our QlikView server to a new one with the focus on switching QlikView and Sense servers one year later, when those QlikView apps will be migrated to Sense.

Migrating from Intel to Intel with dual socket (we have dual socket now) will get an almost 2x increased performance. Migrating from Intel to AMD seems to provide a 3.7x increased performans even though a lower cpu speed.

I'm not sure what to do but sure your comments has been helpfull.

Thank you for your help and the Scalability team.

Pedro

korsikov
Partner - Specialist III
Partner - Specialist III

It's very interesting to know the result of migration. Please, share with us any informations.