Hacker News Reader

Comment by turblety | Hacker News Reader

Do hosting companies no longer give you the speed of a processing core? I couldn't find anywhere that explains how fast 1 of the ARM cores go? Is it 1ghz or 100mhz? Seems like that would be quite important. Seems like 1vCPU is a bit of an arbitary figure. That could be 1 vCPU that goes at 4ghz but then 2 vCPU's that only go at 1ghz. I feel like there's information missing.

View on HN

Just created a new instance. Processor speed is 2297 MHz. lscpu result:

ubuntu@ip-172-31-15-145:~$ lscpu Architecture: aarch64 Byte Order: Little Endian CPU(s): 1 On-line CPU(s) list: 0 Thread(s) per core: 1 Core(s) per socket: 1 Socket(s): 1 NUMA node(s): 1 Vendor ID: ARM Model: 3 Model name: Cortex-A72 Stepping: r0p3 BogoMIPS: 166.66 L1d cache: 32K L1i cache: 48K L2 cache: 2048K

But clock speed doesn't really mean anything these days either. 1Ghz is just as arbitrary - the instruction rate through two CPUs at the same clock rate can vary massively.

So a vCPU is an arbitrary unit but as each is equivalent then they can be benchmarked for the relevant workload.

vCPUs are not at all equivalent across instance types! One vCPU is a single hardware thread on an actual core (ie, two vCPUs is a full core of whatever the underlying machine is if it is an SMT2 core). So the power of a vCPU is quite different across different hardware types.

You might be thinking of ECU (EC2 compute unit), which are intended to be comparable across hardware and are normalized to some old 1.7 GHz CPU that I guess was common in the early days of EC2. Amazon doesn't promote the ECU rating for instance types much, but it's still available if you look.

Just to clarify, if you read below, you'll find that vCPUs on the A1 instances are each physical ARM cores.

Right, I didn't mean to imply otherwise - but I can see how what I wrote could come off that way.

One vCPU is a single hardware thread, which means that on cores without SMT, one vCPU == one core.

Since the A72 doesn't have SMT, you get one core per vCPU. So in that sense, you are getting more bang for your buck if you just count physical cores: twice as many cores per vCPU...

FWIW, I tried to look up the ECU rating of the new ARM instances but they are listed as "NA" in the EC2 console.

GHz doesn't matter anyways. In general nothing matters since threads can be individually throttled without you noticing (apart from declining performance). See my Lightsail review: https://www.karoly.io/amazon-lightsail-review-2018/

That depends on what you're paying for. If you're using Lightsail or one of the t-series EC2 instances you'll definitely get throttled as you don't get dedicated hardware. All other EC2 instance types give you dedicated hardware, on the Intel instances you get one hardware thread per vCPU and on the ARM instances you get a physical core. Those instance types don't throttle.

Challenge accepted ;-) Will benchmark them as well.

Details are there in appropriate pages.

> C5 instances feature the Intel Xeon Platinum 8000 series (Skylake-SP) processor with a sustained all core Turbo CPU clock speed of up to 3.4GHz, and single core turbo up to 3.5 GHz using Intel Turbo Boost Technology. C5 instances offer higher memory to vCPU ratio and deliver 25% improvement in price/performance compared to C4 instances, with certain applications delivering greater than 50% improvement. C5 instances provide support for the new Intel Advanced Vector Extensions 512 (AVX-512) instruction set, offering up to 2x the FLOPS per core per cycle compared to the previous generation C4 instances.

https://aws.amazon.com/ec2/instance-types/c5/

Haven't looked up details for the new stuff yet, might take some time to be up.

With AWS, I found this info from tableau recently:

An AWS vCPU is a single hyperthread of a two-thread Intel Xeon core for M5, M4, C5, C4, R4, and R4 instances. A simple way to think about this is that an AWS vCPU is equal to half a physical core. Therefore, when choosing an Amazon EC2 instance size, you should double number of cores you have purchased or wish to deploy with.

https://onlinehelp.tableau.com/current/server/en-us/ts_aws_v...

Don't know how true it is, but there's no reason for me not to trust what Tableau has to say.

It’s true. Also note that you get a physical core for each vCPU with t2.

I’ve benchmarked c5 and t3 against t2 and have found single thread perf to be better on c5 compared to t2. When loading all vCPUs the performance suffered on the newer instances (each HT performed worse than a single t2), so you would get more bang for your buck on t2. YMMV of course.

Yes T2 performs quite admirably! If your CPU-bound workload doesn't use much memory, smaller sizes of T2 Unlimited can be a much better deal than C4.

T3 gives you both threads of your physical core and commensurably more CPU credits per hour to utilize them. You may notice that even T3.nano offers 2 vCPUs where T2.small has 1 vCPU.

It is in the AWS documentation:

https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instance...

"Each vCPU is a hyperthread of an Intel Xeon CPU core, except for T2 instances."

Most hosting companies do. Cloud hosting (a small subset, as far as I know), generally don't.

If you believe the sales pitch is to help you know that 2 ECU will be twice as fast as 1 ECU (without having to worry about, does this chip has 8MB of L2 cache or 12?).

If you're more cynical, you'll probably think they do it to help hide the fact that cloud computing tends to get trounced badly across all performance aspects (raw cpu, ram, disk io and network io) while being quite a bit more expensive.

> Cloud hosting (a small subset, as far as I know)

Am I living in a bubble, or you are living under a rock?

I'd be surprised if AWS/Google/Azure accounted for more than low digit % of all "hosting".

Last hard numbers I saw, Amazon was the single largest by # of servers, but OVH wasn't too far behind. There's thousands of these worldwide, most older than any cloud provider. Some of the more well known: OVH, Rackspace, Softlayer, Hetzner, Hivelocity, Online.net and on and on.

Add to that people doing their own infrastructure (Apple, Facebook, Google, ...). Companies doing huge colo and on premise (banks, governments). Then consider even larger providers and telcos: AT&T (where WoW was hosted, might still be), Telefonica, Equinix (which I believe AWS was merely a customer of in Singapore at some point), ...

That still leaves VPS, shared hosting, and whatever you want to call DO/Linode/Packet/Vultr.

Reply on news.ycombinator.com