197 points by mike_heffner 9 months ago
It would be nice if AWS could write something official about what they are doing.
I've been noticing major performance changes in our instances and have no idea if it is related to Meltdown or something else.
Google released a blog post specifically on performance: https://blog.google/topics/google-cloud/protecting-our-googl...
It would be nice to have similar transparency from AWS.
Same, I feel it's irresponsible not to provide more details about this issue.
Google likely only wrote a blog post about it because they were able to find a way to brag that there was effectively no performance hit. They have never written any blog posts IIRC explaining bad/unpredictable performance on GCE
Do they not deserve to brag about it? Heck they among others are who found the flaws as well as Broadpwn, Cloudbleed, Heartbleed among others and deserve credit for the work they have done, imo.
Think the constant ragging on Google that appears on HN is becoming a little too much.
It was Google who discovered the vulnerabilities in the first place. How long did they internally wait to give themselves and advantage before disclosing these? I'm honestly asking I don't know.
That's still cool and worth bragging about.
Amazon will probably earn more from autoscaling due to the slowdown.
That's one interesting aspect of these issues and mitigations is that performance really depends on the workload. Just because Google saw little performance impact on their servers, doesn't mean your application won't see. Or because someone said their CPU usage went up 2x doesn't mean it will go up for you.
On an unrelated note, kind of wish Meltdown had been discovered and exposed separately from Spectre. Intel has managed to weasel its way of out of taking responsibility by implying that this is not a bug and all the other CPUs have similar issues. If they had to respond to Meltdown only, it would have made it a bit harder for their PR and legal department to deny the security and performance implications.
Disclosure: I work on Google Cloud.
Just a nit, we said  that not only are our own applications in production doing fine (even against Variant 2) but also we haven’t been inundated with support calls over the last few months while mitigations were silently rolled out at the host kernel and hypervisor layer. So this class of “Hey, my instances are suddenly way slower, I didn’t do anything” isn’t happening on GCE.
That does not mean that if you perform guest OS updates, don’t have PCID enabled, etc. that you won’t see a degradation. We certainly haven’t tried all permutations of all guests, and the kernel patches are still improving. That’s why we’re actively trying to get everyone to rally behind retpoline, KPTI (with PCID enabled), and so on.
> but also we haven’t been inundated with support calls over the last few months
Sorry, I didn't mean to imply that it was Google's customers specifically who should measure and be suspicious. I meant in general, say someone running a service on bare metal outside of any public cloud for example drawing this conclusion - "Google measured the impact of these bugs and mitigations and they didn't see any a significant performance regressions so I probably don't have to worry either".
Also (since you disclosed you work on GCP), I like how Google sponsored Project Zero. Fantastic work and I am sure it will be a great return on that investment. I can certainly see someone thinking about that when deciding to go with GCP vs other solution.
My team saw a 40% CPU usage increase on all of our EC2 instances and even our RDS instances. We were shocked since the media was downplaying the performance impact.
I tried to start a poll but it seems as though my team was just the unlucky one: https://news.ycombinator.com/item?id=16109036
> 40% CPU usage increase
When I read this I thought how weird. It is usually claimed that max performance drop is around 30%. Then I remembered that we are talking in ratios, so a 30% drop in performance(0.7) means a 42% increase in CPU usage(1/0.7=1.42).
How are you trying to measure the performance impact? Are you checking the cloudwatch data or running any specific test?
I am interested to assess the baseline statistics (unpatched so far) and after patching. Any suggestions?
my teams usage is pretty even so the cloudwatch graph is pretty obvious: https://m.imgur.com/a/khGxU
That is the clearest graph ever to show the degradation in performance. How do you get your usage so flat?
Are your ec2 instances PV? They are known to suffer perf issues more than the other type which you can switch to.
Not sure about RDS
Do you have premium support with AWS? Maybe try opening a ticket.
Pardon this likely naive question, but I haven’t seen it addressed yet in all the coverage: what’s the cost in electricity of patching this vulnerability? Does a company like amazon running a massive cloud infrastructure see a non-negligible increase in their cost of doing business?
It may impact the AWS control plane and amazon.com, but as far as AWS services go, it just means customers will be paying for more instances.
But for VM, do customers pay by CPU usage or runtime? I am sure the Netflix of this world optimise their CPU usage but I also suspect the majority of VMs are mostly idle or have little traffic as their task are either intermittent or are sized for peak usage.
Not on AWS but as far as I can tell from their pricing estimator, for 24/7 instances you pay based on time running, not cpu usage.
Can see the appeal of AWS for scaling, redundancy, cross region availability, etc., but for long lived services it seems you pay a high price compared to services like Digital Ocean where the per VM cost is significantly lower.
Technically, you pay for both time (hours on) and CPU Usage (instance tier). Its not like different instance tiers (at least in the same class) use fundamentally more or less powerful processors. They all use the same processors, you just get more or less of it depending on what you pay.
Conceptually it is "pay as you use" by CPU usage, but just rounded into buckets by instance tier.
Of course, there's a lot of underutilization within each bucket, because the granularity isn't per 1% used, but (more or less) per 100% used (aka each core). And also, most applications can't switch instance tiers easily to adapt to demand (though some certainly can).
>>> Its not like different instance tiers (at least in the same class) use fundamentally more or less powerful processors. They all use the same processors, you just get more or less of it depending on what you pay.
There is a variety of CPUs. You can "cat /proc/cpuinfo" to see what you got.
>>> most applications can't switch instance tiers easily to adapt to demand (though some certainly can).
Most applications can and do change. Changing the instance type is just a reboot of the machine.
> There is a variety of CPUs.
There certainly are, but generally they are the same within the same "class" (ie t2, c5, etc). Example:
- C5: 3.0Ghz custom Xeon processors.
- C4: Xeon E5-2666 v3
- C3: Xeon E5-2680 v2
- X1/X1e: Xeon E7-8880 v3
- R4: Xeon E5-2686 v4
- R3: Intel Xeon E5-2670 v2
Etc. The only exception I'm aware of are the burstable instances; AWS's documentation doesn't explicitly guarantee a CPU bin there, it just says "high frequency Xeon".
The difference between an m5.large and an m5.xlarge isn't that you're getting a "faster" processor (meaning higher gigahertz or newer architecture). You're just getting "more" of the same processor (more cores). This is different than on, say, GCP, where you just ask for cores and you can specify if you specifically want a broader generation of Xeon (like Broadwell), but you can't be guaranteed specific chips.
My intention behind saying that most applications can't switch instance tiers is more to point out that most applications aren't prepared to handle node failure, not that there is something intrinsic to the node types which stops them from being able to switch (that'd be much more rare).
Would love to know if anyone else had data on:
* Impact on M5/C5 instances over similar time period, any difference with the Nitro hypervisor?
* Were Dedicated instances (https://aws.amazon.com/ec2/purchasing-options/dedicated-inst...) patched as well?
* Other examples of software that adapted batching performance automatically with increase in call latency.
Not able to answer your questions, but a comment on the article -
>During this same time period, we saw additional CPU increases on our PV instances that had been previously upgraded. This seems to imply some level of HVM patching was occurring on these PV instances around the same time that all pure-HVM instances were patched
This is likely due to Vixen: https://lists.xenproject.org/archives/html/xen-devel/2018-01...
>.... Instead of trying to make a KPTI-like approach work for Xen PV, it seems reasonable to run a copy of Xen within an HVM (or PVH) domU .....
>.... all PV instances in EC2 are using this ....
So the initial bump after the reboot would have been the shim hypervisor which mitigates Vixen. The secondary bump, and bump the native HVM instances saw, would have been the Spectre related stuff.
Based on https://aws.amazon.com/security/security-bulletins/AWS-2018-... - guessing Intel microcode updates
We had a lot of m5 and c5 servers randomly die. It was as if someone was running chaos monkey from Netflix in our VPCs...
Likewise. Can you reach out to me privately? I'd love to have independent corroboration.
Could you send a list of instance IDs and timeframes where you saw this?
Anyone has more info on the performance recovery today? We experienced similar performance issues over the last few days with a seemingly complete recovery today (on a cluster of ~2500 HVM T-1s).
Very curious as to what changed today if performance increased. Some sort of smarter patch? That'd be an amazingly impressive thing to cobble together so quickly.
This is especially interesting for workloads that already ran on >70%
Some stuff won't run in the free tiers anymore and people will have to switch to bigger machines :/
We saw instances which normally kept a healthy stock of CPU credits quickly burn through them and severely degrade in performance thanks to Meltdown :<
Trying to foresee the future...
Could we expect Intel to fix the design flaw^Wfeature so that future server appliance (but also desktop) can run without KPTI while still not being affected by Meltdown ? If so, what timeline could we expect ? Say a year for new CPU designs, plus a year to roll-out new machines in datacenter ?
What I’ve seen is it takes five years to design a CPU from scratch.
I imagine they’ll try and rush this to get it out there as fast as possible (obviously a lot of people would like to buy CPUs they don’t have this issue for security/performance reasons) but it’s going to take a while. I think years is definitely the minimum.
Meltdown is easy enough (relatively) but Spectre is kind of a disaster. What do you do? Does the branch predictor have to start tagging every branch guess with some sort of process ID to prevent one process from messing with another’s predictions? Tag the cache lines instead so even though the data is in cache you can’t see it because YOUR process didn’t pull it in yet? What a mess.
>Does the branch predictor have to start tagging every branch guess with some sort of process ID to prevent one process from messing with another’s predictions?
Its worth pointing out that for their newest designs AMD (and Samsung Exynos) uses the full memory address for branch predictions; no doubt Intel's next design will be doing this.
Ah, that makes since. Sounds like a much less complicated fix than my idea.
Just isolate code on different cores.
"Why I like to run my own hardware for $100, Alex"
You can patch various tiers of servers at your own leisure, depending on threat levels and exposure. Measure the impact, capacity plan, etc. Rather than it being forced on you across all tiers because cloud.
You forgot to type 5 or 6 zeros there.
Granted, that's the bottom of the barrel (single disk, no IPKVM etc.), but $100 keeps you running for over a year. Better servers are easily available as well, usually a couple of times cheaper than AWS.
Is this a US thing? Based on HN only, I'd never know there's anything between the public cloud and racks of own hardware that you have to wire up and maintain.
I have a bunch of quad core 32 GB machines with dual 480GB SSDs for less than $100/month each (and that's a rather expensive provider with great support, you'll cut the price almost in half with e.g. SoYouStart).
Yes, AWS is convenient, but it's far from the only thing in the world.
HN has a lot of professionals. They can't run a business on a refurbished server without ECC and without RAID and without dual power supplies.
Saying that they should run on kumsufi is like explaining to a wholesale company that they should use motorbikes instead of trucks, because motorbikes are cheaper.
Of course. I'm not telling anybody to run on kimsufi, like I'm not running kimsufi myself, but on brand-new Supermicro machines that check all your boxes.
Kimsufi is still dedicated hardware (however low-end) that costs 1 figure instead of 5/6. For normal server-grade hardware it's just 5x cheaper than AWS, not 20x.
Kimsufi and soyoustart are brands of OVH. You don't need to explain again and again that they sell cheaper servers than AWS. Same for super micro. The HN audience know the common suppliers.
Again, you're arguing for motorbikes instead of trucks because they are cheaper. Noone care about that, trucks were not selected based on costs in the first place.
AWS Advocates Eventual Consistency, and I believe offers less than 3 nines guaranteed uptime on many products.
We've been taught to build distributed system with unreliable componens and temporaral JIT eventual consistency.
Of course we can run production cloud-scale operations on unreliable systems. And single power supplies without ECC or RAID is pretty low on my list of things that cause outages. Most big hadoop/cassandra shops are running without raid and without redundant power.
hadoop/cassandra are replicated to multiple nodes by the software. That's a terrible argument for not using any raid in regular setups.
I suspect your and my definition of regular setups is different. My regular setups are stateless and automatically installed and configured.
Lucky you, never having to deal with any database or storage.
Yeah, and AWS has a free tier. $0/month is better than $100/month, right?
HN types are enamored with the scale and size that is necessary to run the massive framework for their blogs.
If you know what you are doing, it is leaps and bounds cheaper to run your own hardware (co-located, rented from soneone else). The only issue is latency on scaling out (hours), but if you are halfway decent with trend lines you can preempt this.
That's not the only issue; there's also a lot of compliance issues that having a hosting company can take care of. There's whole sections of PCI and HIPPA compliance that you can just write off as "not our problem, talk to AWS".
PCI compliance had little to do with where you're hosted, but how you're hosted and what you do with data.
AWS is not a magical PCI compliance button.
Theoretically, it isn't. In actuality, and in our experience, there are a lot of compliance standards where just saying "we're on AWS" gets you 90% of the way toward acceptance. Its mostly buzzword compliance.
That's just not true. Most of PCI compliance is documenting practices and operations. Have you even completed all various levels of compliance? Or have you just done one level?
And what if the delta between your upper and lower daily “trend lines” is measured in millions of requests per hour? Per second? We can leave off weekly/seasonal trends for now, and keep it nice and easy for you.
The utter lack of imagination that I see on HN when people are judging others’ technical decisions is kind of hilarious.
It was a Jeopardy! reference. Sorry if it didn't come through.
It's very easy for a single admin with a single machine to provide 4 nines to a small group with a small load. But this usually scales exponentially.
Just because you can't do it doesn't make it impossible. A small team of 4-5 good ops people can scale a network of many thousand nodes, petabytes of storage, and terabits of network throughput.
Tech stuff isn't hard. The biggest problem is the lack of capacity planning and project communication in tech today. Nimble startup is a euphemism for pure anarchy and chaos. No one wants to plan anything any more.
The same infra could be managed by a single guy if it were in the cloud.
No, because that "guy" would want a vacation and would suddenly become the most important person in the company. You're running teams of people, not cogs. Once you realize that, you'll also realize that cloud services are like 5-6x the total cost.
Cloud services are not efficient or cheap, they're convenient for the unprepared and haphazard management style that VC's love.
You should review the difference between CapEx and OpEx, and specifically focus on why companies might prefer to pay “5-6x the total cost” when that cost comes in the form of OpEx.
You may still disagree in the end, but you should come to the conclusion that it is not categorically stupid for some companies to choose to pay much more over time so that they can pay a lot less right now.
Only by abstracting away hundreds of teams that manage the actual hardware.
Over a 15 year time scale, there is no way AWS will remain competitive with GCP.
Why is that ? They both belongs to the top cloud providers today, both seems to invest highly on their R&D. I don't see anything obvious on what could happen in the future.
Google’s network engineering is better.
Is there an option of AWS dedicated instances without these patches? I thought all these new vulnerabilities are only really dangerous in shared environments.
> I thought all these new vulnerabilities are only really dangerous in shared environments.
You're not the first I see saying that, here and on other sites, and this is absolutely wrong.
Shared environments like clouds were singled out because not only were they impacted the worst security wise, they were also going to suffer the most from the fixes.
But even if you only have a regular normal happy server or computer for you alone, remote code execution vulnerabilities aren't unheard of; one your application (be it your own, or a specific one you use, or one of the bazillion stuff running on your system as part of the OS) gets broken and you're a free target. Or anything with a proper scripting surface.
If you're system isn't protected, and any of the application you run has a major security hole, everything will be at risk.
Guess humanity lost 30% of its computing power