Date: Wed, 8 Nov 2017 00:59:23 +0000 From: Rafal Lukawiecki <raf@rafal.net> To: Colin Percival <cperciva@tarsnap.com> Cc: Pete Wright <pete@nomadlogic.org>, freebsd-cloud@freebsd.org Subject: Re: AWS c5.class unsupported? Message-ID: <F511BDC0-65FE-420E-9CDF-25D39B2B28C5@rafal.net> In-Reply-To: <0100015f99189134-61e4c63e-1f35-4854-ad9c-b4b62bb084a0-000000@email.amazonses.com> References: <22772d45-812b-4592-87b7-7e9bafcc86f7@email.android.com> <53A5CAA0-B37C-4BAE-A42A-338C04D9612E@rafal.net> <0100015f99189134-61e4c63e-1f35-4854-ad9c-b4b62bb084a0-000000@email.amazonses.com>
next in thread | previous in thread | raw e-mail | index | archive | help
> On 8 Nov 2017, at 00:47, Colin Percival <cperciva@tarsnap.com> wrote: > > Actually no -- the problem I'm aware of has nothing to do with user-data > scripts and should not affect instance launching. I wouldn't be surprised > if some instance launches are failing due to capacity issues though -- EC2 > often has such problems when they launch a new instance family. > > The problem I'm tracking with C5 relates to disk hotplug: If you add an EBS > volume, FreeBSD will not notice it until you run 'devctl rescan pci0', and > if you remove an EBS volume FreeBSD will panic. Warner is working on fixing > the "hotunplug" path in the nvme driver in order to correct the panic; we > were hoping to have it fixed in HEAD before C5 was launched, but I thought > we had a few weeks left. > > If you have a good test case for user-data script problems, please let me > know (off-list). Given how much new hardware and software there is in the C5 > instances I wouldn't be surprised if there are a few bugs lurking, and I'm > sure my contacts at EC2 will be happy to receive reports. > > (FWIW, C5 not being available in the Marketplace images is probably just a > case of the Marketplace taking a while to get up to speed; I marked those > images as "make these available on all new instances". But I'll probably go > in and disable C5 for now due to the disk attach/detach issues.) > > -- > Colin Percival > Security Officer Emeritus, FreeBSD | The power to serve > Founder, Tarsnap | www.tarsnap.com | Online backups for the truly paranoid It was odd. Your cloudconfig seemed not to run *some* of the scripts (or *none* of the scripts) that it got from the tar instance user-data. It could have been related to a fluke co-incidence. During the project development, I have, so far, launched 150+ t2.medium and c4.large and I have never had this issue. Today I thought of testing c5.large but I have only launched 12 of them. I had this problem in 1 case out of the 12. On a side note, the performance our optimised web heads running FreeBSD 11.1 + PHP 7 + Drupal 7 got from c5.large was on-par or a little worse than from t2.medium in a short but intensive load test. I was surprised, so perhaps there are some other issues lurking beneath. I was also hoping the slowish EFS access would be faster, but it is no different at the low-end of EFS storage it seems. Of course, t2 is limited to bursts, so this is not a fair comparison to c5, but I was still hoping to see better results even in a 5 min load. More testing needed, results inconclusive, I suppose. I am on a deadline to complete the project this week, so I won’t go back to c5 testing but happy to do that in about 5 weeks from now. Rafal
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?F511BDC0-65FE-420E-9CDF-25D39B2B28C5>