New AWS I3 Instances for Demanding, I/O Intensive Applications | Hacker News Reader

New AWS I3 Instances for Demanding, I/O Intensive Applications

aws.amazon.com

84 pointsmanigandham28 comments

Wow. This is a massive price drop from the I2 instances... so massive that I question whether I'm reading it correctly:

    i2.xlarge: 4 vCPU / 14 ECU, 30.5 GB RAM, 1 x 800 SSD, $0.853 per Hour

    i3.xlarge: 4 vCPU / 13 ECU, 30.5 GB RAM, 1 x 950 SSD, $0.312 per Hour

Slightly less CPU, slightly more SSD, 63% cheaper. Or for another comparison:

    r3.xlarge: 4 vCPU / 13 ECU, 30.5 GB RAM, 1 x  80 SSD, $0.333 per Hour

r3, unlike i2, is listed as "current generation", but also seems strictly worse than i3.

I interviewed with Amazon several months ago (and turned down the offer, so I have no affiliation); as part of the casual conversation with the people there, it was mentioned to me that the company hadn't offered instances optimized for IO up until that point due to other priorities but had been recently focusing on it with the I tentions of releasing something like this soon.

I was under an NDA on the topic and so never mentioned it; now that it's been released, i am free to say that this has been worked on for a while and that these sorts of gains in performance were not only expected but a goal.

(No disclaimer: I didn't take the job and none of my work or personal projects run on AWS.)

I don't think it takes any special NDA protected knowledge to say that a new instance type was "worked on a for a while", given how rarely they release new instance types (the i2 family was announced in Dec 2013). Likewise, the i* family is their "I/O intensive workload" family, so again, it's clear that gains in I/O performance was a goal when designing i3's, the i2 family was definitely showing its age, even compared to a modern laptop.

The newer R4:

    r4.xlarge: 4 vCPU / 13.5 ECU, 30.5 GB RAM,    EBS only, $0.266 per Hour

Good point: if you don't need any local storage at all, then r4s are a bit cheaper. Still, an extra $576 / year (the price difference between i3.xlarge and r4.xlarge) for 950 GB of SSD really seems quite cheap, especially as that's the on-demand price.

Once I found out about NVMe providing up to 2x-10x better performance than regular SSDs, I immediately started wondering when Cloud/VPS providers would offer something based on that technology. I think that before U.2 (or M.2?) was a thing, they were stuck with really expensive PCIE cards. Now I am guessing that something like U.2/M.2 NVMe subsystems can make the price more practical.

But back in September I asked on the Linode forum and also on the Uservoice for Digital Ocean about NVMe storage. Basically I got crickets on Digital Ocean, and "don't be an idiot" from one Linode user. https://digitalocean.uservoice.com/forums/136585-digitalocea...

https://forum.linode.com/viewtopic.php?f=7&t=14058&sid=4a574...

But I think that as more and more people get computers with M.2 built in, they will eventually start to wonder why their PC's disk is 2x-10x faster than their server's disk.

There are a couple of little providers like UltraVPS, ControlVM, and Gignode that at least say they have NVMe available. I have been thinking about UltraVPS because it looks like a deal.. of course don't know if it really does perform or is reliable.

Basically my thing is that from my experience the faster IOPS/sequential read translates into less database tuning or zero database tuning depending on the application I am coding.

Just to throw another wrinkle in here, I feel like pretty soon the inexpensive VPS/cloud servers will not only have to move to NVMe, but within a few years the expectation may be to include some kind of GPU hardware somehow as a standard thing. Just because neural network stuff seems so important for so many applications these days, and unless we are going full peer-to-peer internet or something I am going to need that on most servers going forward pretty soon.

Are these comparable to Google's Local NVMe SSD? On Google, the drives alone are as expensive as these i3 instances.

375GB*3 drives on GCE will cost you 0.3431/hr over a month (incl. auto discount), which is already more than an i3.xlarge, and you still have to pay for CPU/RAM.

Just from the IOPS listed in the text, they're about half the speed per drive as ours (3.3M/8 => 400k) but as before they let you stripe together way more than we do. I'll be curious to see what happens with wear and garbage collection... (we're pretty conservative, these things come in powers of two and the resulting output size of bytes means they're more aggressive or optimistic).

Definitely an awesome update!

Disclosure: I work on Google Cloud.

16 cores + 122GB + 3.8TB for less than $1k/month is impressive.

On GCP: 16 cores + 104GB + 8x375GB (for 3TB total) = $1470/month or $1226/month with continuous use discount.

As I've replies in the other comment, but will make more explicit here: I think our drives have roughly 2x as many IOPS/GB (just using their marketing statement in the blog post). There's lots of reasons you might prefer one over the other (and AWS lets you stripe way more together), but just comparing GB isn't a full picture (otherwise PD-HDD or EBS Magnetic or even GCS/S3 is way more awesome...).

Disclosure: I work on Google Cloud.

An equivalent server to the first one mentioned, if rented monthly from e.g. wholesaleinternet.com, is ... $209/month and includes 33TB transfer.

No wonder that AWS is so profitable.

Maybe less so when you do the math and find you can buy a similar machine for something like 4-5k, and use it for 3 or so years?

The on-prem vs cloud argument is sufficiently tired. Obviously you should do the math and figure out what's best for you, but that's not really related to AWS releasing new VM types.

If you're running on AWS for whatever reason, i3 VMs are great news.

If you're using it for three years, then you'd take an RI and it'd be 60% of the cost. You also don't have to worry about failed parts, hardware monitoring, racking, power, etc, plus the skills required to manage same.

Yes it is more expensive than buying your own hardware, but I thought we all understood at this point that AWS does a little more than just give you an empty server and no additional services.

You can't really do that kind of math without using AWS incorrectly. The reality is that you won't run the instance on AWS for 3 years. You spin it up when you need it, and you kill it when you don't.

There will be some instances that you'll run 24/7, but this shouldn't be the case for your entire fleet if you're on AWS.

I seriously doubt you could do that. Those SSD's are ~1.5k each right alone and the power/infrastructure to keep it running reliably with a fat connection to the internet isn't cheap.

If only they would have waited for Ryzen :-P

I'd like to know what kind of motherboards they are using that support up to 8 NVMe drives!

Nothing special; each Xeon socket can provide 5 PCIe x8 slots. Don't try this in a 1U though.

Yeah, sorry, brainfart. I have only ever used NVMe drives in M.2 so I completely spaced that you can put 'em in PCIe slots too (which is probably more popular, actually?).

Thinking about it, that (2 x NVMe in PCIe) will probably be the first upgrade I do to my new workstation in the future. It looks like those only use four lanes each and I've got enough still available (80 total) to support way more of these drives than I expect to ever need on my desk.

ETA: PCIe v3.0 x16 == 15.75 GB/s. I expect we'll see, at some point, PCIe cards/adapters that support up to 4 NVMe drives (in one slot). Hell, they might exist already.

Ask and you shall receive: https://www.servethehome.com/the-dell-4x-m-2-pcie-x16-versio...

And those NVMe devices don't necessarily get dedicated access, they can definitely be over provisioned for access to PCIe

Because of the heat or something else?

A 1U physically doesn't have room for 8 PCIe cards, although come to think of it you can fit 10 U.2 SDDs.

Wow if these run as fast as they say one could replace tens of GCP machines with a single i3.xlarge, or you could go with a 2-4xlarge just for giggles and still come in far under budget.

Huh? The i3.xlarge is clearly two slices of a full 2T device. The performance numbers in the text are for eight of those. While it'd be great if we had 8 of ours available instead of just two, I don't understand how you believe you'd replace even more than one GCE VM with a single i3.xlarge.

Disclosure: I work on Google Cloud.

10's of machines? i3's are definitely dense in performance but GCE can be stacked pretty well too.

The biggest issue beyond pricing though is that GCE has a limited total disk bandwidth regardless of the number of local SSDs attached.

We should probably make our documentation more clear. You should see fairly linear scaling as you go from 1x375 up to 8x375. Is that not the case? (We can only go to 8x375 today, but that's not "regardless" of how many drives you attach, that's "Can only attach so much").

Disclosure: I work on Google Cloud.

Reply on news.ycombinator.com