Open Stack

We ($work) have been researching this topic for the past few weeks and I
wanted to give an update on what we've found.

First, we've found that both Rackspace and Azure advocate the use of
RAID'ing block storage volumes from within an instance for both performance
and resilience [1][2][3]. I only mention this to add to the earlier Amazon
AWS information and not to imply that more people should share this view.

Second, we discovered virtio-scsi [4]. By adding the following properties
to an image, the disks will now appear as SCSI disks, including the more
common /dev/sdx naming:

hw_disk_bus_model=virtio-scsi
hw_scsi_model=virtio-scsi
hw_disk_bus=scsi

What's notable is that, in our testing, ZFS pools and Gluster replicas are
more likely to see the volume disconnect/fail with virtio-scsi. mdadm has
always been fairly dependable, so there hasn't been a change there. We're
still testing, but virtio-scsi looks promising.

1:
https://support.rackspace.com/how-to/configuring-a-software-raid-on-a-linux-general-purpose-cloud-server/
2: https://support.rackspace.com/how-to/cloud-block-storage-faq/
3:
https://azure.microsoft.com/en-us/documentation/articles/virtual-machines-linux-configure-raid/
4: https://wiki.openstack.org/wiki/LibvirtVirtioScsi

On Mon, Feb 8, 2016 at 7:18 PM, Joe Topjian <joe at topjian.net> wrote:

> Yep. Don't get me wrong -- I agree 100% with everything you've said
> throughout this thread. Applications that have native replication are
> awesome. Swift is crazy awesome. :)
>
> I understand that some may see the use of mdadm, Cinder-assisted
> replication, etc as supporting "pet" environments, and I agree to some
> extent. But I do think there are applicable use-cases where those services
> could be very helpful.
>
> As one example, I know of large cloud-based environments which handle very
> large data sets and are entirely stood up through configuration management
> systems. However, due to the sheer size of data being handled, rebuilding
> or resyncing a portion of the environment could take hours. Failing over to
> a replicated volume is instant.In addition, being able to both stripe and
> replicate goes a very long way in making the most out of commodity block
> storage environments (for example, avoiding packing problems and such).
>
> Should these types of applications be reading / writing directly to Swift,
> HDFS, or handling replication themselves? Sure, in a perfect world. Does
> Gluster fill all gaps I've mentioned? Kind of.
>
> I guess I'm just trying to survey the options available for applications
> and environments that would otherwise be very flexible and resilient if it
> wasn't for their awkward use of storage. :)
>
> On Mon, Feb 8, 2016 at 6:18 PM, Robert Starmer <robert at kumul.us> wrote:
>
>> Besides, wouldn't it be better to actually do application layer backup
>> restore, or application level distribution for replication?  That
>> architecture at least let's the application determine and deal with corrupt
>> data transmission rather than the DRBD like model where you corrupt one
>> data-set, you corrupt them all...
>>
>> Hence my comment about having some form of object storage (SWIFT is
>> perhaps even a good example of this architeccture, the proxy replicates,
>> checks MD5, etc. to verify good data, rather than just replicating blocks
>> of data).
>>
>>
>>
>> On Mon, Feb 8, 2016 at 7:15 PM, Robert Starmer <robert at kumul.us> wrote:
>>
>>> I have not run into anyone replicating volumes or creating redundancy at
>>> the VM level (beyond, as you point out, HDFS, etc.).
>>>
>>> R
>>>
>>> On Mon, Feb 8, 2016 at 6:54 PM, Joe Topjian <joe at topjian.net> wrote:
>>>
>>>> This is a great conversation and I really appreciate everyone's input.
>>>> Though, I agree, we wandered off the original question and that's my fault
>>>> for mentioning various storage backends.
>>>>
>>>> For the sake of conversation, let's just say the user has no knowledge
>>>> of the underlying storage technology. They're presented with a Block
>>>> Storage service and the rest is up to them. What known, working options
>>>> does the user have to build their own block storage resilience? (Ignoring
>>>> "obvious" solutions where the application has native replication, such as
>>>> Galera, elasticsearch, etc)
>>>>
>>>> I have seen references to Cinder supporting replication, but I'm not
>>>> able to find a lot of information about it. The support matrix[1] lists
>>>> very few drivers that actually implement replication -- is this true or is
>>>> there a trove of replication docs that I just haven't been able to find?
>>>>
>>>> Amazon AWS publishes instructions on how to use mdadm with EBS[2]. One
>>>> might interpret that to mean mdadm is a supported solution within EC2 based
>>>> instances.
>>>>
>>>> There are also references to DRBD and EC2, though I could not find
>>>> anything as "official" as mdadm and EC2.
>>>>
>>>> Does anyone have experience (or know users) doing either? (specifically
>>>> with libvirt/KVM, but I'd be curious to know in general)
>>>>
>>>> Or is it more advisable to create multiple instances where data is
>>>> replicated instance-to-instance rather than a single instance with multiple
>>>> volumes and have data replicated volume-to-volume (by way of a single
>>>> instance)? And if so, why? Is a lack of stable volume-to-volume replication
>>>> a limitation of certain hypervisors?
>>>>
>>>> Or has this area just not been explored in depth within OpenStack
>>>> environments yet?
>>>>
>>>> 1: https://wiki.openstack.org/wiki/CinderSupportMatrix
>>>> 2: http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/raid-config.html
>>>>
>>>>
>>>> On Mon, Feb 8, 2016 at 4:10 PM, Robert Starmer <robert at kumul.us> wrote:
>>>>
>>>>> I'm not against Ceph, but even 2 machines (and really 2 machines with
>>>>> enough storage to be meaningful, e.g. not the all blade environments I've
>>>>> built some o7k  systems on) may not be available for storage, so there are
>>>>> cases where that's not necessarily the solution. I built resiliency in one
>>>>> environment with a 2 node controller/Glance/db system with Gluster, which
>>>>> enabled enough middleware resiliency to meet the customers recovery
>>>>> expectations. Regardless, even with a cattle application model, the
>>>>> infrastructure middleware still needs to be able to provide some level of
>>>>> resiliency.
>>>>>
>>>>> But we've kind-of wandered off of the original question. I think that
>>>>> to bring this back on topic, I think users can build resilience in their
>>>>> own storage construction, but I still think there are use cases where the
>>>>> middleware either needs to use it's own resiliency layer, and/or may end up
>>>>> providing it for the end user.
>>>>>
>>>>> R
>>>>>
>>>>> On Mon, Feb 8, 2016 at 3:51 PM, Fox, Kevin M <Kevin.Fox at pnnl.gov>
>>>>> wrote:
>>>>>
>>>>>> We've used ceph to address the storage requirement in small clouds
>>>>>> pretty well. it works pretty well with only two storage nodes with
>>>>>> replication set to 2, and because of the radosgw, you can share your small
>>>>>> amount of storage between the object store and the block store avoiding the
>>>>>> need to overprovision swift-only or cinder-only to handle usage unknowns.
>>>>>> Its just one pool of storage.
>>>>>>
>>>>>> Your right, using lvm is like telling your users, don't do pets, but
>>>>>> then having pets at the heart of your system. when you loose one, you loose
>>>>>> a lot. With a small ceph, you can take out one of the nodes, burn it to the
>>>>>> ground and put it back, and it just works. No pets.
>>>>>>
>>>>>> Do consider ceph for the small use case.
>>>>>>
>>>>>> Thanks,
>>>>>> Kevin
>>>>>>
>>>>>> ------------------------------
>>>>>> *From:* Robert Starmer [robert at kumul.us]
>>>>>> *Sent:* Monday, February 08, 2016 1:30 PM
>>>>>> *To:* Ned Rhudy
>>>>>> *Cc:* OpenStack Operators
>>>>>>
>>>>>> *Subject:* Re: [Openstack-operators] RAID / stripe block storage
>>>>>> volumes
>>>>>>
>>>>>> Ned's model is the model I meant by "multiple underlying storage
>>>>>> services".  Most of the systems I've built are LV/LVM only,  a few added
>>>>>> Ceph as an alternative/live-migration option, and one where we used Gluster
>>>>>> due to size.  Note that the environments I have worked with in general are
>>>>>> small (~20 compute), so huge Ceph environments aren't common.  I am also
>>>>>> working on a project where the storage backend is entirely NFS...
>>>>>>
>>>>>> And I think users are more and more educated to assume that there is
>>>>>> nothing guaranteed.  There is the realization, at least for a good set of
>>>>>> the customers I've worked with (and I try to educate the non-believers),
>>>>>> that the way you get best effect from a system like OpenStack is to
>>>>>> consider everything disposable. The one gap I've seen is that there are
>>>>>> plenty of folks who don't deploy SWIFT, and without some form of object
>>>>>> store, there's still the question of where you place your datasets so that
>>>>>> they can be quickly recovered (and how do you keep them up to date if you
>>>>>> do have one).  With VMs, there's the concept that you can recover quickly
>>>>>> because the "dataset" e.g. your OS, is already there for you, and in plenty
>>>>>> of small environments, that's only as true as the glance repository (guess
>>>>>> what's usually backing that when there's no SWIFT around...).
>>>>>>
>>>>>> So I see the issue as a holistic one. How do you show operators/users
>>>>>> that they should consider everything disposable if we only look at the
>>>>>> current running instance as the "thing"   Somewhere you still likely need
>>>>>> some form of distributed resilience (and yes, I can see using the
>>>>>> distributed Canonical, Centos, RedHat, Fedora, Debian, etc. mirrors as your
>>>>>> distributed Image backup but what about the database content, etc.).
>>>>>>
>>>>>> Robert
>>>>>>
>>>>>> On Mon, Feb 8, 2016 at 1:44 PM, Ned Rhudy (BLOOMBERG/ 731 LEX) <
>>>>>> erhudy at bloomberg.net> wrote:
>>>>>>
>>>>>>> In our environments, we offer two types of storage. Tenants can
>>>>>>> either use Ceph/RBD and trade speed/latency for reliability and protection
>>>>>>> against physical disk failures, or they can launch instances that are
>>>>>>> realized as LVs on an LVM VG that we create on top of a RAID 0 spanning all
>>>>>>> but the OS disk on the hypervisor. This lets the users elect to go all-in
>>>>>>> on speed and sacrifice reliability for applications where replication/HA is
>>>>>>> handled at the app level, if the data on the instance is sourced from
>>>>>>> elsewhere, or if they just don't care much about the data.
>>>>>>>
>>>>>>> There are some further changes to our approach that we would like to
>>>>>>> make down the road, but in general our users seem to like the current
>>>>>>> system and being able to forgo reliability or speed as their circumstances
>>>>>>> demand.
>>>>>>>
>>>>>>> From: joe at topjian.net
>>>>>>> Subject: Re: [Openstack-operators] RAID / stripe block storage
>>>>>>> volumes
>>>>>>>
>>>>>>> Hi Robert,
>>>>>>>
>>>>>>> Can you elaborate on "multiple underlying storage services"?
>>>>>>>
>>>>>>> The reason I asked the initial question is because historically
>>>>>>> we've made our block storage service resilient to failure. Historically we
>>>>>>> also made our compute environment resilient to failure, too, but over time,
>>>>>>> we've seen users become more educated to cope with compute failure. As a
>>>>>>> result, we've been able to become more lenient with regard to building
>>>>>>> resilient compute environments.
>>>>>>>
>>>>>>> We've been discussing how possible it would be to translate that
>>>>>>> same idea to block storage. Rather than have a large HA storage cluster
>>>>>>> (whether Ceph, Gluster, NetApp, etc), is it possible to offer simple single
>>>>>>> LVM volume servers and push the failure handling on to the user?
>>>>>>>
>>>>>>> Of course, this doesn't work for all types of use cases and
>>>>>>> environments. We still have projects which require the cloud to own most
>>>>>>> responsibility for failure than the users.
>>>>>>>
>>>>>>> But for environments were we offer general purpose / best effort
>>>>>>> compute and storage, what methods are available to help the user be
>>>>>>> resilient to block storage failures?
>>>>>>>
>>>>>>> Joe
>>>>>>>
>>>>>>> On Mon, Feb 8, 2016 at 12:09 PM, Robert Starmer <robert at kumul.us>
>>>>>>> wrote:
>>>>>>>
>>>>>>>> I've always recommended providing multiple underlying storage
>>>>>>>> services to provide this rather than adding the overhead to the VM.  So,
>>>>>>>> not in any of my systems or any I've worked with.
>>>>>>>>
>>>>>>>> R
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> On Fri, Feb 5, 2016 at 5:56 PM, Joe Topjian <joe at topjian.net>
>>>>>>>> wrote:
>>>>>>>>
>>>>>>>>> Hello,
>>>>>>>>>
>>>>>>>>> Does anyone have users RAID'ing or striping multiple block storage
>>>>>>>>> volumes from within an instance?
>>>>>>>>>
>>>>>>>>> If so, what was the experience? Good, bad, possible but with
>>>>>>>>> caveats?
>>>>>>>>>
>>>>>>>>> Thanks,
>>>>>>>>> Joe
>>>>>>>>>
>>>>>>>>> _______________________________________________
>>>>>>>>> OpenStack-operators mailing list
>>>>>>>>> OpenStack-operators at lists.openstack.org
>>>>>>>>>
>>>>>>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-operators
>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> OpenStack-operators mailing listOpenStack-operators at lists.openstack.orghttp://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-operators
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> OpenStack-operators mailing list
>>>>>>> OpenStack-operators at lists.openstack.org
>>>>>>>
>>>>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-operators
>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> OpenStack-operators mailing list
>>>>> OpenStack-operators at lists.openstack.org
>>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-operators
>>>>>
>>>>>
>>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-operators/attachments/20160306/45288982/attachment.html>