StorageSummit

several questions today on the reliability of flash

The crowd in NYC was highly virtualized - well over half - folks in DC are mum (bunch of NSA guys???)

What has been the impact of virtualization on storage?

are you getting into any cloud discussions? Do they see virtualization as a first step to cloud?

I never thought that virtualization was necessary for cloud, but probably helps

@stu - cloud is obviously big in the gov't - seems like the fervor around virtualization is definitely more tepid rel to a few yrs ago

@cmosoares I know some guys that don't want a virtualization layer - they want bare metal - even in the cloud believe it or not. I think sometimes virtualization in the cloud benefits the service provider

the low hanging fruit of virtualization has been picked - opportunity is in mission critical applications plus the necessary transformations in storage (SDS) and networking (SDN)

Do you think virtualization has pretty much reached the saturation point?

@stu I agree - now the opportunity is in fine-tuning the infrastructure to handle the virtualization and to speed the applications

@cmosoares I don't think it's saturated but it's going to take longer to attack the hard to move apps. The big question is how the public cloud will impact all this

@oraclehardware impact of virtualization on storage was initially it broke storage. Then massive integration efforts put a band-aid on the problem and now the industry is trying to make storage "invisible" through the sw-defined theme

@chuckhollis has been on a tear writing about SW-defined storage and vsan lately - ServerSAN is another wildcard in that mix as is OpenStack cc: @stu

which apps are typically virtualized and why?

I think the greatest fear in clouds is the "noisy neighbor" problem. If the underlying system can't handle a lot of simultaneous users, then you need to isolate to separate hardware. Not very efficient.

Yes Bob, it's a big issue - Security in the cloud...many people think it's a "do-over"

I think the contention problem is a tougher nut to crack than the security problem -- especially in private clouds.

What about segmentation to optimize based on type of workload? compute, io, transaction profile, etc? Can this be done dynmically?

@JeffFrick That's the challenge, Jeff. It's all of the above. EMC said as much in their VNX2 launch, but we're already there. You need lots of cache and lots of threading, so you can handle lots of VMs without needing to know in advance what's coming.

Not all architectures are created equal: ZS3 is superior for virtualization - SSG-Now Report: http://www.oracle.com/us/products/servers-storage/storage/nas/snapshop-report-analyst-paper-2008431.pdf

no!

ZS3 can support as many as 10,000 VMs per system; EMC announced support for 6,600 for VNX 8000

a write to flash first architecture should deal with the IO blender problem better

ZS3 designed for highly virtualized data centers – Taneja Group: Top Performance on Mixed workloads, Unbeatable for Oracle Databases http://www.oracle.com/us/products/servers-storage/storage/nas/taneja-group-zs3-analyst-paper-2008432.pdf

A couple of Wikibon reports point to hybrid, flash-first architectures http://bit.ly/1icydbB http://bit.ly/NYPhn6

Definitely not. Most storage vendors (other than Oracle) don't have enough cache/CPU or an SMP OS to be able to handle random, heavy VM I/O to begin with...not to mention VM-level analytics to be able to resolve VM impacts.

I would think dtrace would be an effective tool in finding hot spots in, for example, a vmware situation - the io blender problem makes performance mgt a 'science project' for many customers

There are many factors that effect performance, architecture, integration levels but also (very importantly) the application itself and how well it's designed

DTrace part of Oracle ZS3 data services provide highly granular VM-level view of the data stream, so admins can quickly spot the VM(s) that's causing the performance bottleneck. See here http://bit.ly/1m84Rcn

@dvellante Large, consolidated virtual environments show off dtrace better than any other workload. Based on the conversations we're having, people aren't thrilled with the analytics on competing storage systems, which are cobbled together.

security is always a priority at scale, but we have had relatively few issues at this stage. As we increase in size these problems will become more relevant and laying proper groundwork to insure we can handle these issues is a priority.

In respect to the noisy neighbor problem in particular, we have definitely noticed AWS stealing time from some smaller machines we overload occasionally, which can be really annoying, but seems effective at making everyone play nice

@drew_bratcher AWS is all about low cost. The next phase here is selecting gear (and yes, we think we have it at Oracle) that can do a lot of consolidation without sacrificing service quality.

We had a Peer Incite a while back on the Noisy Neighbor problem - here's the outcome of that http://wikibon.org/wiki/v/Quieting_Noisy_Neighbors_In_Cloud_Services

Interesting report, Dave. Another way to handle noisy neighbor is what we did inside Oracle IT. They consolidate 2,300 VMs on a single Oracle ZFS storage box that can handle triple that. This leaves a ton of room for bursts and it's easy.

maybe you should ask what "crapplications" have you virtualized :-)

It used to be virtualization as a test-case for non-mission critical apps, now it's definitely more mainstream

@oraclehardware-> When does the conversation flip from "What should we visualize" to "What can't be virtualized?"

@JeffFrick from a technical perspective there should no longer be any barriers to virtualizing any application. That being said, bare metal isn't going away and there are alternatives (such as PaaS) to server virtualization.

@KerstinWoods it's funny to see the industry go full circle - IBM mainframes were/are all virtualized

What's the optimal storage architecture to support highly virtualized environments?

All that consolidation of workloads at the server level definitely has a storage impact - performance, boot storms, lack of VM visibility, distance from the application.

Look at the 'mess' the storage industry had to deal with because of server virtualization http://wikibon.org/wiki/v/VSphere_V5_Storage_Integration_Points

What's best? all-flash? Hybrid?

@cmosoares ZS3 is superior for virtualization - @SSG-Now Report: http://www.oracle.com/us/products/servers-storage/storage/nas/snapshop-report-analyst-paper-2008431.pdf

@cmosoares Hybrid is ideal - flash performance with disk cost-efficiency, within that category ZS3 takes the lead @DGIG
https://go.oracle.com/LP=1427?elqCampaignId=3753

A couple of Wikibon reports point to hybrid, flash-first architectures http://bit.ly/1icydbB http://bit.ly/NYPhn6

here's Wikibon's report on multi-hypervisors from last summer http://wikibon.org/wiki/v/VMware_Dominant_in_Multi-Hypervisor_Data_Centers

@dfloyer put together this research on multi-hypervisors http://wikibon.org/w/images/2/22/DistServerswithHypervisors.png