Seven days and one server: how we tested AMD-based machines

We had one server based on a new AMD processor, a bunch of benchmarks that we wanted to run, and a week to test the machine under real load. Not that this was enough time for a close acquaintance with a fresh piece of hardware, but when you constantly test new items and customize services, it is difficult to stop.

What tests did we manage to run, in which tasks the “newbie” outplayed our workhorse from the flagship pool, and where he lost – we tell under the cut.

Introducing the test subjects

We have been waiting for the opportunity to take a server with the latest generation of AMD processors for testing. Recently, such an opportunity presented itself to us. Meet the Dell EMC PowerEdge R6515 Powered by AMD EPYC 7742 Processor.

Key CPU characteristics:

  • 64 cores;

  • 128 threads;

  • base frequency 2.25GHz;

  • maximum frequency per core (boost) up to 3.4GHz;

  • L1 cache – 4MiB;

  • L2 cache – 32MiB;

  • L3 cache – 256MiB.

Agree, it looks promising. Recently, we have already taken for a test servers with processors from AMD of the previous generation. The result, alas, was not impressive – approximately at the level of Intel five years ago.

And then enthusiastic articles about new CPUs began to appear on the Internet. This inflamed skepticism in us – there is a lot of talk, but no one has objective test indicators. Everyone wrote that there was more performance, reduced power consumption. That Dell EMC PowerEdge servers with AMD processor are suitable for demanding resource-intensive applications and cloud services (they use AMD EPYC chips with 8 to 64 cores and support high-speed PCIe 4.0). Therefore, we decided to get fresh processors at any cost and run at least the main set of tests on them. Since we are engaged in virtualization – roughly speaking, we give the cores to customers – we were interested in how the CPU would behave under load.

We will compare this server with our current “workhorse” – Dell Poweredge R740 with two Intel Xeon Gold 6254 on board. We have been actively using these servers for about a year now. The processors there are excellent and suitable for any task. Except, perhaps, 1C. Higher frequency CPUs are needed here. For 1C we use Intel Xeon Gold 6244. Here wrote how the Gilev tests were carried out on them.

Pool of test tasks

Our standard testing procedure is at two levels:

  • a series of tests on the server itself, “through the eyes of a provider”;

  • testing from virtual machines located on the server – “through the eyes of the client”.

Unfortunately, the process of debugging the server inside our infrastructure took some of our precious time, and we only managed to run a few sysbench tests. Nevertheless, the results are quite interesting.

Test suite

Test

cmdline

sysbench, max prime, one core

taskset -c 0 sysbench –test = cpu –cpu-max-prime = 20000 run

sysbench, max prime, all cores

sysbench –test = cpu –cpu-max-prime = 100000 –num-threads = 8 run

sysbench, oltp-mysql, i thread

sysbench –test = oltp –db-driver = mysql –mysql-db = test –mysql-user = root –mysql-socket = / var / lib / mysql / mysql.sock –mysql-table-engine = innodb –max-requests = 0 –oltp-table-size = 1000000 –max-time = 300 –num-threads = $ i run

All tests were run on a virtual machine with 8 vCPUs and 32Gb RAM.

Results: CROC standard pool

For starters, let’s look at the numbers from the reference Dell EMC PowerEdge R740:

Test

Result

sysbench, max prime, one core

total time: 19.1545s

sysbench, max prime, all cores

total time: 22.1102s

sysbench, oltp-mysql, 1 thread

828.69 tr. per sec.

sysbench, oltp-mysql, 2 threads

1605.72 tr. per sec.

sysbench, oltp-mysql, 4 threads

2992.22 tr. per sec.

sysbench, oltp-mysql, 8 threads

5927.20 tr. per sec.

Results: AMD based server

The results of the tested Dell R6515 with AMD EPYC 7742:

Test

Result

sysbench, max prime, one core

total time: 15.6657s

sysbench, max prime, all cores

total time: 18.9329s

sysbench, oltp-mysql, 1 thread

1023.46 tr. per sec.

sysbench, oltp-mysql, 2 threads

1709.39 tr. per sec.

sysbench, oltp-mysql, 4 threads

3231.34 tr. per sec.

sysbench, oltp-mysql, 8 threads

4533.65 tr. per sec.

As you can see from the results, the virtual machine on the R6515 performed better than on the R740, except for the OLTP test for 8 threads – here the advantage remained with the reference machine. It was OLTP testing that opened a small pitfall: in threads 1, 2 and 4, the performance is excellent, and in 8 threads the processor is already “buried”.

It is still difficult to say why this is happening. To understand the processor and learn how to work with it, one week is clearly not enough. I would like to carry out additional tests in various variations: for example, “drive” a VM with 4 cores into four threads. This would allow us to understand the peculiarities of the CPU.

It would be very interesting to run the beloved Gilev test and compare the results. Alas, we had a limited time, so we tested the most basic.

conclusions

Let’s discuss the characteristics of each of our subjects and try to summarize.

The Dell EMC R6515 server with AMD EPYC 7742 is definitely interesting. It has a high density of cores per unit, good performance and a slightly lower cost compared to Dell from our flagship pool: a benefit of about 40% per vCPU and 20% based on actual performance (price / performance).

The disadvantages include high heat generation, but it all depends on the air conditioning system in your data center.

The result is a powerful harvester for large volumetric tasks, capable of digesting serious analytical calculations, CRM systems, databases or many virtual machines without any problems. In terms of cost, this one processor costs less than two Intels from our standard pool. And the server itself is cheaper than the currently used R740. The result is higher performance for less money.

My colleagues and I came to the conclusion that AMD’s processor turned out to be really good. It is perfect for implementing typical scenarios. Perhaps AMD EPYC 7742 will appear in our cloud – we are constantly customizing services, so such an animal can be very useful.

Of course, we cannot recommend them unequivocally. The choice of hardware is based on tasks. Test different options and choose what works best for you.

By tradition, we are waiting for your comments. If you have any questions, we will be happy to answer them.