> From: [email protected] [mailto:[email protected]]
> On Behalf Of Josh Smift
> 
> which leads me to
> speculate that other than memory, they may be pretty equivalent in terms
> of their network performance and reliability.

Responding by quoting myself:

> On the "Instance Types" page they describe clearly:
> http://aws.amazon.com/ec2/instance-types/
> 
> Type  Networking Performance
> t2.Micro      Low to Moderate
> t2.Small      Low to Moderate
> m3.Medium     Moderate
> m3.xlarge     High
> And so on...
> 
> The point being, they do in fact prioritize network traffic in favor of larger
> more expensive instances.

I'm going to expand upon this by saying, "Low to Moderate" is vague and 
undefined.  The actual observed performance varies - sometimes 20-30 Mbit/sec 
and 20ms latency, sometimes 15 KB/sec and >500ms latency.

I'll give them the fact that "Low to Moderate" is vague, but it wasn't always 
like this.  For years I was comfortable with the (low) performance I saw out of 
tiny systems, for doing simple little webpage servers (such as nedharvey.com).  
It's only in the last several months that it's gotten this bad.


> which lends some
> credence to
> the idea that the network unreliabilty that Ned is complaining about is
> *not* in fact normal T2 low-end-ish-ness, but an actual problem, perhaps
> (as someone suggested) with the underlying hardware that that particular
> instance is running on.)

A few comments.  -1-  The normal, steady state performance has degraded 
significantly over the last year.  So that's one thing.  The "normal" T2 
low-end-ish-ness has gotten worse.  To the point where I wouldn't call it 
"normal" anymore.  Even a tiny instance should not timeout serving simple 
webpages, and should be capable of exceeding the internet performance 
expectations of 20 years ago.  15 KBps and 700ms latency is IMHO outside what a 
normal person would call "normal" level of "Low to Moderate" network 
performance.  "Low to Moderate" is not synonymous with "Dead in the water."

-2-  During the last month, when we were seeing hundreds of alerts per day, 
that was *not* normal.  It was definitely caused by an underlying problem.  We 
have solid reason to believe the specific underlying problem was Amazon's need 
to reboot machines for the xen bug, and they were supposed to finish by Oct 1, 
but in fact, our last problem day was Oct 6, and our machines were not supposed 
to be affected (because our machines did not need to reboot).  However, during 
all that maintenance, it became clear to see, that all the additional workload 
in the system adversely affected tiny instances *far* beyond what anyone would 
call "normal" low-end-ish-ness.  Fortunately this maintenance period seems to 
be over now, but it shows us that the tiny machines will get the boot as soon 
as anything higher priority asks for scheduling.

-3-  It's not *specific* hardware having a problem.  Over the last several 
months, the general low-end-ish-ness problem has been observed on several tiny 
(and I think small) instances, in different companies where I work, being 
created, destroyed, stopped and rebooted.

The "normal" low-end-ish-ness that I declare IMHO is not what a normal person 
would consider "normal," is wildly variable, by a factor of 1,000.  (Sometimes 
MBit/sec, sometimes KBit/sec).  This enormous variability despite having idle 
CPU cycles and memory, is indicative of network (and/or network storage) 
prioritization.
_______________________________________________
Tech mailing list
[email protected]
https://lists.lopsa.org/cgi-bin/mailman/listinfo/tech
This list provided by the League of Professional System Administrators
 http://lopsa.org/

Reply via email to