#serversan - CrowdChat

Scalable Infrastructure

Chat with thought leaders, analysts, and customers on new server/storage architecture

#serversan @stu @deepstoragenet @gunnarwb storage

LeaderBoard

#ServerSanHyperconverged InfrastructureSharing new Wikibon Server SAN research and interactive Q&A

Stream Ended

Establishing a secure connection

What do you folks think about data locality? If the same group of servers are doing compute and storage should the data for a VM be on the server where that VM is running? or is that adding complexity?

[+] Show Hidden Comments

David Nicholson

It is a trade off. Parallelism can be powerful. Compromise is an ability to create "domains" and "pick your media/connectivity".

Generally speaking, data is sticker, and moving the compute or having it co-resident is an advantage that #serversan has. added benefits are also reduced latency

Howard Marks some of the products do locality and others don't. Which has lower latency and net traffic? We'll have to see.

@dfloyer -> Part of David Floyer's Mega Data Center story.

John F. Kim Not to be confused with BMW's Mega City (i3) story

Why would you want that in a virtualized server San fabric? What about VM migrations to other physical hosts?

Ideally you'd like it close but the devil will be in the detail. What happens when you want to replace a node(s)....

We go back and forth. For some apps (e.g. Hadoop), the benefit could be worthwhile. For others, not so clear. And then there is the complexity factor.

Jesse St. Laurent

It is almost always easier to move the VM than the data. A few GB of vRAM vs. 100s of GB of data

then the system needs to be smart enough to handle that in a resilient elegant manner

Jesse St. Laurent

The admin should not need to know or care. The system should simply make this decision.

http://wikibon.org/wiki/v/Keep_Your_Own_Data_Close_and_Your_Competitors_Data_Closer

Keep Your Own Data Close And Your Competitors Data Closer - Wikibon
Why Register?

Enrico Signoretti19

@tier1storage Latency could be a serious problem, in certain situations Infiniband could be better. but I'm not sure that enterprises like infiniband anymore...

[+] Show Hidden Comments

Jesse St. Laurent

I do not see customers with IB infrastructure. Even small customers are moving to 10GbE

John F. Kim If they have certain scale-out storage products or database appliances, they already have InfiniBand in the data center.

Enrico Signoretti Yep! 10GbE is getting cheaper and cheaper. Servers are shipped with 10GbE ports now at the same price (or very similar price).

Enrico Signoretti I agree that 10GbE is the present and 40GbE is the future. Eth won years ago.

Larry loves infinband! :-)

Enrico Signoretti larry loves catamarans too but they are not in the DCs. In any case, they use infiniband primarly in the backend (which I like a lot!)

John F. Kim I wonder if Larry's yacht has its own InfiniBand network. If not, will look into fixing that.

Gabriel Chapman

not enough diversity in vendors in the IB space to make it competitive for average customer, Eth is cheap in comparison

Jeff Frick -> "Bacon is for Closers" - I like that.

Enterprises don't like IB. But today's 10/40/100Gbs Ethernet is under 2us latency so good enough. Especially with RoCE

Jesse St. Laurent Competing with "good enough" is very difficult. Competing with Ethernet is a recipe for going of business

David Nicholson

If #serversan is about driving out cost and complexity, then IB needs to be cheap and easy to play.

Rodger Burkley19

The issue with SFW RAID and other software defined storage models was always one of less performance...till now it seems

[+] Show Hidden Comments

Yes, general purpose CPUs have gotten so powerful this has become a non-issue

Barry Ader Exactly!!. It is clearly one of the reasons why storage services can be delivered in the server. "The world is flat, again"!

John F. Kim Agree! Nearly all proprietary arrays use software on general purpose CPU, only a few models still have proprietary silicon.

Barry Ader http://scaleio.com/blog/entry/the-world-is-flat

David Nicholson Fast, cheap server interconnects help also. No #serversan without the SAN part. :-)

There have been 2 big reasons Software storage was slow. 1 - Not enough disks/controller 2 - Synchronous mirroring across slow Ethernet. SSD fixes first, 10Gbps Ethernet the second

John F. Kim Or 40Gb Ethernet and 56Gb InfiniBand for sync replication (shameless plug for my company's wares)

Jesse St. Laurent I completely agree. None of this is possible without 10GbE and flash

Jesse St. Laurent

The data redundancy is now being delivered in software, now the question is how to deliver high performance data efficiency (dedupe, compression, optimization). We always want something more that what software offers.

So many cores to play with there is no need for dedicated ASICs etc like 3PAR. New platforms can handle RAID and data optimization like compression and dedupe in SW. Aided by SSD.

Dave Vellante will be interesting to watch - obviously the chaps at HP/3PAR don't agree

Storage Godfather (HPEStorageGuy) David Scott has said that he has the engineering team justify keeping the ASIC with each new release of HP 3PAR ASIC. They're a smart team and will make the right choices.

Software is that it is easier to change than firmware and an ASIC. Speed of entry and future development wins unless there is a huge advantage from a hardware based implementation. The incumbents defend with their technical base and acquire the new tech.

[+] Show Hidden Comments

But anything can be committed to silicon...including "programmable" ASICs...:-)

Amen to that. Doing things in software makes it WAY easier to patch things in the field and speeds development.

Jesse St. Laurent

With a good HW team, HW can deliver innovation very quickly. It also keeps the workload off the Intel CPU and allows that CPU to be used for the customers business apps

Jesse St. Laurent

Every CPU cycle the storage consumes also consumes part of a vSphere (or other hypervisor) license

Depends on the hardware. I like the SW only approach, but SimpliVity is doing some cool stuff with hardware that helps them enable their value prop.

I think in general we've ignored the fact that on a Server SAN the server CPU being used for storage is going to mean there's less CPU for VMs. If running ServerSAN means I need 36 servers instead of 32 is that better than 32 and a storage system or 3.

[+] Show Hidden Comments

Don't forget the consider the soft costs that come from the simplicity in the solution. That can't be overstated enough.

Ditto on the simplicity and consistency angle -- comes out all the time in customer discussions.

Jesse St. Laurent

I spoke with a customer recently who told me the OPEX cost of an additional platform was 10x the cast of buying more Lego blocks

Howard Marks Question there is what the additional platform was. Best in class storage (TinTri, Tegile, Nimble) a lot easier than Clariion

David Nicholson

Haven't ignored it. Cores become expensive real estate when an Oracle license moves in. The ability to fence off cores from #serversan duty can be beneficial.

And, don't forget that these are building block solutions with small increments. Other solutions eventually require a forklift at some point in many cases.

Howard Marks True enough but again there have to be times when best of breed beats the really good swiss army knife

Scott D. Lowe Howard - agreed. Server SAN is just another option in the arsenal.

Jesse St. Laurent @deepstoragenet Sometimes theSwiss Army knife is the "best of breed" :)

Chuck Hollis and that's what I see customers trying to figure out now -- where does it fit in the landscape?

Perhaps we should be thinking in terms of CPU cores rather than servers...plus with software defined storage you can manage and balance the system dynamically, right?

Gunnar Berger11

Question for VARs: I see the ServerSAN as an enabler to a simpler datacenter. For VARs this could be bad as complexity increases service fees. I often wonder how the mighty dollar works against a simpler solution

[+] Show Hidden Comments

Speaking as a former CIO: Simplicity good. Increased service fees bad. :-)

Gunnar Berger Agreed, end user customers would like this, which is why I asked for a VAR. Most of my experience is as a VAR and when a sales guy makes more selling X, they will try and sell it.

VARs like both more service fees from complexity *and* simple easy solutions to increase deal velocity. Okay, maybe not always the same VAR liking both.

the value chain is always flexible - if VAR is making $ for not adding value they need to reevaluate skill set and solution development.

Jesse St. Laurent I agree. The channel is in the process of reinventing itself right now. The best ones embrace compelling technology and take the opportunity to move their value up the stack.

Honestly, I think VARs in some cases are going to need to reinvent. Between this and things like O365 (i.e. MS reducing shared profits), the model is not turning in their favor.

I see this a huge opportunity for VARS to add value in services esp in light of hybrid cloud and big data plans coming together

Dave Vellante11

Does anyone have an opinion on how/if the security model is impacted by server-san -- what are the implications?

[+] Show Hidden Comments

Enrico Signoretti

Good point, but it should not happen with the correct design. isolation is not difficult to achieve.

I see potential for improving security in some cases. Storage is certainly more isolated in these models (ie. accessible only to the hosts and not available from elsewhere)

But that, too, is dependent on the solution.

Security w/software defined architectures provide more flexibility, enabling isolation and multi tenancy, both important considerations for large enterprises and Service Providers

Scott D. Lowe The software nature seems like it could be more resilient to security issues, too... more easily patchable.

David Nicholson

Software Defined Security will address any issues. I am only half-joking. We have more abstraction/virtualization to deal with. Threats that dwell need to be ferreted out with analytics. Not good enough to build a wall. P or V.

Any features still missing compared to traditional arrays? distance replication / snapshots etc. reliant on higher level sw for this?

[+] Show Hidden Comments

Great question chris - have to believe the stack is nowhere near as mature as a 20YO array's function - but would think innovation will happen fast

Jesse St. Laurent If you start with a clean slate and build the right core, it is amazing how much easier it is to bring new features to market. The 20YO+ products have 20 years of baggage.

of course. But many of the solutions are adding those features very quickly.

David Nicholson

As a general rul #serversan lags behind the 25 + years of development that is part of modern arrays

interesting point is - Server SAN should NOT necessarily require the same amount of functionality as a traditional array, rather new architectures and sw can do it differently

Scott D. Lowe Agreed. But features are always being added.

Jesse St. Laurent

It depends on the platform. Not in the @SimpliVityCorp platform. Inline dedupe, compression, replication, backup to the cloud, VM centric policy, etc.

David Nicholson

But the gap will close quickly

Since these products are all in their early releases some have any given feature but few have all

Most of these features will be added to Server SAN over time if not already there. Maybe biggest missing link is ability to add app servers without running Server SAN software on them?

Jesse St. Laurent Some of the Server SAN products offer that today

Jesse St. Laurent It is a critical feature you you want to truly decouple the scalability of CPU, memory, storage performance, and storage capacity.

David Nicholson Actually some allow that flexibility. Sometimes "patch tuesday" or "Oracle licensing" make this a requirement.

Which of course begs the question where should services like snapshots and replication be provided. If hypervisors had good snaps and replication it would be a real argument

Chris Wilson Exactly :) Just trying to get my head around how this would all hang together compared to the solutions we deploy today

We called it Software Defined Storage but too many vendors deiced that fit what they did too

[+] Show Hidden Comments

What makes SW Def Storage?

Howard Marks To me Software defined storage HAS to be software, If you wrap it in tin it's an array,. Then it has to turn RAW storage (probably DAS) into shared

nuance that I'd make is that SDS can still be a separate storage array which Server SAN is combined compute (usable) and storage

John F. Kim Yes Server SAN definitely combines app-usable compute and storage

David Nicholson I would further nuance by saying the "option" to fully converge compute and storage = #serversan

Sylvie Otten (Sollod If we can converge server SAN SW functionality with compute and network, we'll have the basis for an entirely new and more powerful model http://chucksblog.emc.com/chucks_blog/2014/01/of-server-sans-and-sds.html

Of Server SANs and SDS - Chuck's Blog
It was great to read Wikibon's first take on a new market segment -- "server SANs" -- written by my colleague Stu Miniman (@stu). Of all the various industry analysis available, the Wikibon content regularly does it for me, at least when it comes to ...

Wasn't sure at first about "Server SAN" but best alternative I had before was "Scale-out server local storage" which is too long to Tweet