Recommendation on Servers, I like Vultr but what about others?

Introduction

Someone was asking about Vultr HF and Hetzner, but was curious if there was other options. So I responded with this.

Response

Just a couple of observations.

One rare issue with Vultr is that if you size up your instance and there is no capacity on the current node, they will move it to another node. You only find out after your servers have been down for hours and you reach out to support. Depending on your instance size, it could take 1-2 hours, and your instance is down during that time. They need to improve on this, or maybe they have.

Vultr provides the typical CPUs most other VPS providers are using. Intel Sky/Cascade Lake with AMD Rome and Milan. Which are decent, but not the “best” https://managingwp.io/wp-moar-speed/vultr-plans-and-processors/

You can see a list of sysbench single-core test for a handful of providers at https://www.vpsbenchmarks.com/labs/cpus_by_singlethreaded_sysbench_perf just be aware

• People will call this site questionable, but sysbench and Geekbench are standard tools.
• There isn’t every provider and plan available which sucks.
• Not all tests are recent; make sure to check the date.
• Most providers will provision new accounts on the least used instance first, so your first months will be faster and then degrade.
• Most providers will oversell, there’s a healthy middle but some don’t adhere to this and you won’t find out until your node is full or a noisy neighbour starts causing CPU steal.
• Make sure to watch CPU steal.
• Vultr has had issues with noisy neighbors, and you can request to be moved. You can monitor this with monit or Netdata.

As for the timeout issue you’ve been experiencing. Do you have snapshots set up? This could be interfering, you also need to make sure your OS has the proper guest tools to support snapshots. But this should already be in-place if you used a Vultr installation image.
It’s only really an issue with a custom install. It’s also could be related to the nodes snapshotting for all VMs is causing the problem. Regardless Reach out to support, this is not acceptiable, and they should correct it. It might be a noisy neighbour, so check htop for cpu steal.

Some other notes.

For the providers that charge per hour, you can spin up a VM run a sysbench and figure out what CPU they’re providing for a couple of pennys. Then shute them down when the test is finished. It’s a good way to test these providers and find out what CPUs they’re using.

With the help of a few others, we went out and looked at bare-metal options from different providers and collected some data that I’ve been updating. Here’s the AirTable with the list of various providers and the processors they offer https://managingwp.io/projects/wp-moar-speed/

Why does the artable matter? WordPress loves high-frequency and fast single-core processors. Unfortunately, there aren’t many server processors available with high single-core speeds and a large core count due to science. And providers need to maximize rack space. So they need to jam as many cores into 1U as possible. So, they opt for higher core count and lower frequency. So you’ll 32-64 core count CPU’s in a providers node, and also dual CPU to get to 128 cores. But with 3Ghz to 3.8Ghz and low single core performance.

Granted 3Ghz to 3.8Ghz is considered high-frequency. However, there are desktop processors that start at 4.7Ghz and boost to 5.6Ghz; boosting is a whole other topic with much higher single-core performance.

There are a few VPS providers out there that provide Ryzen Desktop VPS instances. But they don’t own their data centers or hardware, so they’re not ideal. However, AMD has recently released AMD Epyc 4xxx series processors, and so more VPS providers should be embracing these and using what are called blade enclosures to house up to 32 instances of these within 4U’s.

But the cost will still be higher than going bare-metal. With the likes of OVH and Hetzner, the cost to performance for bare-metal is huge, with the added benefit of no neighbours to steal CPU. Some might argue, that it’s more management and hardware involved which leads to failures. The biggest issue in the past was spinning disks, which are now replaced with NVMe drives and hardware has become more reliable. So bare-metal have a low failure rate. There is also tools like Netdata to help with monitoring hardware issues.

Vultr is great, but the cost to performance is just so bad it’s hard to not look at alternatives.