GeekBench Results for Cloud Servers
-
@pete-s said in GeekBench Results for Cloud Servers:
@scottalanmiller said in GeekBench Results for Cloud Servers:
@pete-s said in GeekBench Results for Cloud Servers:
@scottalanmiller said in GeekBench Results for Cloud Servers:
@pete-s said in GeekBench Results for Cloud Servers:
@scottalanmiller said in GeekBench Results for Cloud Servers:
@pete-s said in GeekBench Results for Cloud Servers:
.... except that Vultr uses workstation hardware, at least on one machine.
Intel classifies that model as Server...
https://ark.intel.com/products/130046/Intel-Xeon-W-2175-Processor-19_25M-Cache-2_50-GHz
No, it's W-series and W stands for Workstation:
New Intel Xeon W Processors Deliver Optimized Performance for Mainstream Workstation Professionals
It says that on the group page, but look at the specific page for that processor. Says server right on it, and on its parent page.
Yes, but that's the vertical segment. There's no segment for workstation, hence the reason it doesn't say workstation. Intel only have a few segments: desktop, embedded, mobile and server.
But have a look at Dell or HPE. You'll find it next to impossible to find a rackserver with a workstation CPU like W-2175. But you'll find it in plenty of workstations (Dell Precision, HP Z). That should tell you something.
PS. Pricing is not a lot different between the workstation and server CPUs.
Comparable to the 14 core W-2175 @ 2.5GHz is 14 core Intel Xeon Gold 6132 @ 2.6 GHz.
Rec. price: $1947 versus $2111.But Dell and HPE don't focus on single CPU servers, which has long been something that they lacked. Their models are long build around the higher cost dual CPU configurations. Which would explain that more than anything.
When Intel's single socket workstation CPU is the same price as one multi-socket server CPU, why should they bother? They have single socket AMD Epyc servers which makes sense when it comes to price and performance.
Performance perhaps. It's tuned differently. And $200 isn't the same, that's a bit of money in a cloud decision matrix.
-
AMD Epyc is more recent. I'd like to see more of that, too. But you can't just move from Intel to AMD in a single environment. Very few providers have moved to Epyc because one reason or another, mostly from established norms, I would imagine.
-
I was planning to get an AMD Epyc server for testing but as you mentioned you can't do live migration between Intel and AMD hosts. So I figured I need at least two nodes, preferably more. With enough memory and storage it adds up so from a budget perspective it didn't make sense. In the end I got a couple of refurbished servers instead with Intel CPUs.
-
@pete-s said in GeekBench Results for Cloud Servers:
I was planning to get an AMD Epyc server for testing but as you mentioned you can't do live migration between Intel and AMD hosts. So I figured I need at least two nodes, preferably more. With enough memory and storage it adds up so from a budget perspective it didn't make sense. In the end I got a couple of refurbished servers instead with Intel CPUs.
I think that's why a lot of clouds haven't gone to it, yet. But I see new deployments trying it out.
-
@scottalanmiller said in GeekBench Results for Cloud Servers:
AMD Epyc is more recent. I'd like to see more of that, too. But you can't just move from Intel to AMD in a single environment. Very few providers have moved to Epyc because one reason or another, mostly from established norms, I would imagine.
Intel's partner eco-system is one of the best out there IMNSHO. We get great support from them with advanced warranty replacement for any product found to be defective.
When we ventured into AMD a while back to see what was up their eco-system was too fragmented with too much room for willy wagging between hardware vendors whose components were in the servers we were building.
Most, if not all, of Microsoft's A Series VMs back in the day were built on AMD Opteron. When running in-guest performance tests for CPU on the equivalent Intel and AMD instance (Core/vRAM) the Intel absolutely killed AMD for performance.
I'd like to see the same side-by-side comparisons of the current AMD EPYC and Intel Scalable. AMD most certainly has the PCIe Gen3 limitations somewhat mitigated by adding all of those lanes, but that becomes a somewhat moot point if the CPU is not as efficient as the equivalent Intel product.
-
@pete-s said in GeekBench Results for Cloud Servers:
disk
One cannot test for Disk numbers in any real way at least not reliably.
All of the hosting environments we set up have a set of storage QoS policies that are configured according to their customer facing plans. Azure does this and I'm sure that AWS and others also do the same.
EDIT: Better said: One cannot test for in-guest Disk numbers …
-
-
@emad-r not bad, a little better than average (median).
-
Topic like this should be pinned and done every year or so. No one keeps alot of track about this.
What else we can test efficiently...
RAM Speed ? but who cares about this.
How we can test VM Cloud Download/Upload
-
@emad-r said in GeekBench Results for Cloud Servers:
Topic like this should be pinned and done every year or so. No one keeps alot of track about this.
Its' true. A regular survey of what is out there would be very nice. Or even a weekly one!
-
This is my Vultr 5$ 1 vCPU, 1GB RAM.
Yours has an extra 200 Mhz due to bigger package.
I think Vultr is the best for CPU and IO speed from what I feel
-
@emad-r yeah, if you go up to 2GB of RAM, you also move up to the SkyLake processors.
-
Amazing that Vultr's low end systems also outperform everyone else! That's crazy how big the gap is between them and basically everyone else.
-
@emad-r said in GeekBench Results for Cloud Servers:
I think Vultr is the best for CPU and IO speed from what I feel
I agree, and Price too
-
@emad-r said in GeekBench Results for Cloud Servers:
Topic like this should be pinned and done every year or so. No one keeps alot of track about this.
What else we can test efficiently...
RAM Speed ? but who cares about this.
How we can test VM Cloud Download/Upload
Thought about Network testing and came up with this:
iperf3 -c bouygues.iperf.fr -t 30 = upload iperf3 -Rc bouygues.iperf.fr -t 30 = download
-
@pete-s said in GeekBench Results for Cloud Servers:
As you said it might be the best cost performance ratio for cloud deployment.
ehhhhhhh. It's the best cost for selling someone a modernish core for hosting that lacks any real scaling issues, or licensing per core/socket concerns.
-
@emad-r said in GeekBench Results for Cloud Servers:
Thought about Network testing and came up with this:
That's just a testing of peering.
-
@storageninja said in GeekBench Results for Cloud Servers:
@emad-r said in GeekBench Results for Cloud Servers:
Thought about Network testing and came up with this:
That's just a testing of peering.
Help us out then, I think we at ML need to come up with standard for testing Cloud.
And how is it peering ? cause peering means:
the exchange of data directly between Internet service providers, rather than via the Internet.I am using the same separate server of iperf on both nodes and both nodes are located in the same geolocation.
-
@emad-r said in GeekBench Results for Cloud Servers:
@storageninja said in GeekBench Results for Cloud Servers:
@emad-r said in GeekBench Results for Cloud Servers:
Thought about Network testing and came up with this:
That's just a testing of peering.
Help us out then, I think we at ML need to come up with standard for testing Cloud.
And how is it peering ? cause peering means:
the exchange of data directly between Internet service providers, rather than via the Internet.I am using the same separate server of iperf on both nodes and both nodes are located in the same geolocation.
The point is you cannot test certain things. Such as network throughput and disk speed when it comes to cloud.
You can test throughput to your office. But again based on peering agreements, you could get get hugely varying results compared to someone on another ISP to the same cloud service.
-
Right, the base problem is that because you can't control everything, you really can't test cloud servers in a super meaningful way.
To some extent, this goes far beyond cloud servers and kind of applies to all systems that are shared or cross the Internet. It's a more general problem.
