4 ms·
Always nice to see new options, but I wish Amazon hadn't dropped the ephemeral disk. Quite apart from the issue of surviving EBS failures, ephemeral disks are
by cperciva 14y ago
Always nice to see new options, but I wish Amazon hadn't dropped the ephemeral disk. Quite apart from the issue of surviving EBS failures, ephemeral disks are wonderful for swap space -- since the lifetime of swap exactly matches the lifetime of ephemeral disks. I even have a startup script in FreeBSD which autoconfigures swap space striped across a few GB of ephemeral disks.
- JoachimSchipper 14y agoStupid question: is that in the article, or..? I seem to have missed it.
- cperciva 14y agoNot in the article. On the EC2 website: http://aws.amazon.com/ec2/instance-types/ http://aws.amazon.com/ec2/instance-types/
- nilsbunger 14y agoSpecifically, the new instances (m3.*) are only available with ebs. There is probably a technical reason for this - eg maybe they move these bigger instances around to different servers so they can't be on local disk? Would love to know why...
- cperciva 14y agoI'm guessing that Amazon has wanted to phase out ephemeral disk on non-HPC instances for a long time, since it costs them money and most people don't use it; and the availability of EBS-optimized instances and provisioned-IOPS disks was what they needed to satisfy the people who insisted that ephemeral disk was the only way to get reliable I/O performance. And I'm sympathetic to that argument where large amounts of disk are concerned... but I do wish we could have a small amount of ephemeral disk to be used for swap space.
- kelnos 14y agoI think there's a reliability argument to be made as well. Most (all?) of the EC2 outages and partial outages over the past year or so have had their root cause somewhere in EBS. It's also why I hesitate to touch ELBs or RDS, since they both rely on EBS.
- raverbashing 14y agoMy guess is that ephemeral disks were dropped because of too many people losing data on them. Not because of a fault, but because they don't know instance disks are ephemeral. Permanent storage as a default is a saner default as well.
- cynicalkane 14y agoI think it's moot. For heavy users, the use case for ephemeral disk is better served by their I/O instances. For light users, I think Amazon is betting that EBS is 'stable enough' by now. If you really don't like EBS, there's always the old instances, Amazon RDS, Amazon DynamoDB...
- paulsutter 14y agoAll of the disk performance benchmarks (iops, throughout, buffered vs unbuffered) I ran on ephemeral disks matched exactly the performance of ebs. That could be a coincidence that my tests ran only on machines without local disk, but more likely ephemeral disk was really just EBS in most cases. So it always seemed like a silly or even misleading product to me.
- cperciva 14y agoI've never seen ephemeral disks die during an EBS outage. Also, they don't cost anything for I/Os.
- jspthrowaway 14y agoEphemeral is definitely not EBS, as its I/O does not contribute to saturating an instance's network connection (I've tested that). EBS has a ceiling for throughput (the instance's network throughput) and competes with your other network traffic, which is interesting for database workloads for obvious reasons. In addition, every single benchmark I've seen run reports ephemeral behaving very differently from EBS. As ephemeral disappears when you stop an instance (presumably, when the system is given the opportunity to relocate your instance to another machine), I've always suspected that ephemeral storage was part of local disks on the virtual host's chassis -- as opposed to a SAN or other kind of network-attached storage. As such, you probably end up with the same unpredictable performance you find in all virtualized resources, since it is very unlikely that Amazon is giving you your own storage. Before benchmarking ephemeral storage you have to pre-warm it, which might have contributed to your findings. Ephemeral is worlds better than EBS, particularly in outage scenarios; if I could convince everybody on planet Earth to stop using EBS, it would be a noble cause.
- paulsutter 14y agoHmm, I have been told by Amazon and confirmed in my testing that EBS is immune to the network throttle (unlike machismo to machine tcP, s3, etc)
- 14y ago
- newhouseb 14y agoYeah, applications servers (read: web front-ends) can easily run entirely from memory, so it's annoying to have the rest of the OS dependent on EBS when it really needn't be. (Although to be honest I'm not aware how easy or hard it is to pull out the hard disk from a booted instance without customizing the kernel).
- saurik 14y agoIt is quite possible that this is specifically why these new instance types can exist: as a lot of users don't use the ephemeral drives at all, even for swap, you can now build a new set of servers that are just CPUs and network cards, getting rid of the large disk drive; some tradeoff has to be made to increase some of the stats but not the price, and saying it is a shame that they didn't include everything seems to then miss the point.
- csense 14y ago> some tradeoff has to be made to increase some of the stats but not the price I disagree. According to Wikipedia, it's been six years since EC2 launched. Since 2006-equivalent hardware is much cheaper in 2012 due to Moore's Law, (or alternatively the specs you can buy with a given number of dollars in 2012 is much better than it was in 2006), and since the original EC2 machines have probably been in service long enough to generate sufficient revenue to recover their upfront investment and make a profit, it wouldn't be unreasonable for Amazon to start retiring those old machines and use the better price/performance point of new hardware as a business justification to give customers some combination of price breaks and spec upgrades.
- saurik 14y agoThat would make sense across the board; in this case, you have to just compare the price point between different current offerings; these new instance types are similarly priced yet with higher specs in some areas, and I will argue that it is a tradeoff against the disks. If you feel like Amazon should be cheaper overall due to Moore's law (and remember: they have lowered their prices numerous times since they launched on everything from hardware to bandwidth), then the price of the things we are comparing to would also be cheaper (as these are all just virtualized subsets of larger machines).