xflash

Next Gen Server Flash
Analysts discuss the impact of flash on hyperscale for the enterprise.
Top Votes
John Furrier
What is the difference and why should any care about All Flash Arrays vs Hybrid Arrays? CIOs want to know this. I get asked this a ton.. thoughts?
[+] Show Hidden Comments
[-]
Dave Vellante
all flash is the race car whereas hyrids are the souped up production vehicle
Stuart Miniman I think your analogy works for adding flash to older arrays (tier-1 hybrid), but not newer unified hybrids which have been built for flash+disk w higher % of flash
David Floyer If you have to go to disk, you have high IO variance. Database performance is determined by locking rate, and low IO latency AND low IO RT variance are key metrics.
[-]
David Floyer
All flash arrays have lower consistent latency - the variance of the IO in a well designed all-flash array is always very low. In hybrid systems some of the IO comes from disk - by design the variance is high. High activity DBs love low variance.
Dave Vellante so is it correct that hybrids can go after a broader set of use cases and if so which is a bigger market?
David Floyer The broader market at the moment is supporting existing applications. The higher value market of the future is supporting emerging applications.
[-]
Jean-Luc Chatelain
At the end of the day, the how is less relevant that the what. CIO will care for technology which gives best $/GB/s and best $/IO/s because workloads have varied profiles
David Floyer Hi Jean-Luc - That is true for many applications. For some very high business value application low latency will be an imperative. Designing a flat database where all functions can operate in parallel has very high business value.
[-]
Jean-Luc Chatelain
@dvellante I respectfully disagree on your car analogy. It is only a matter of the design of the hybrid. Posche Cayenne is a 4WD with the heart of a sports car and same can be done for hybrid arrays.
Dave Vellante I would agree re: the Cayenne - I would argue the F1 is purpose built for a more narrow use case (i.e. all flash for DB)
[-]
Jean-Luc Chatelain
@dvellante Dloyer calls its xflash, I call it IME (Infinite Memory Engine) but yes in the end it is about giving a memory semantic and hide the ugliness of disk protocols and limitations
Dave Vellante so how will this infinite memory pool change application design?
[-]
Jean-Luc Chatelain
@dvellante It does not change designs but it evolves with the app. If app is uses file/block semantic, it looks like file/block but if app is true in-memory then it uses a memory semantic mmap() for example
Dave Vellante what % of today's apps are written with a memory-dominated semantic in your est. and how will this change over time?
David Floyer
However, for future designs off applications where analytics are integrated with transactions it will be an essential ingredient to have very low latency IO
[+] Show Hidden Comments
[-]
Stuart Miniman
there are a number of solutions (@pernixdata @qlogic @proximaldata) that network flash, how is this different?
[-]
Ed Walsh
Do you anticipate future application designs will intelligently leverage a relatively small pool of high performance storage and bulk tier of storage with rich data services?
Dave Vellante that's a great question - will those data services by SW-defined and independent of media or will they be "superglued" to the device?
David Floyer Ed - absolutely - a key requirement is the ability to ensure that active data is in the flash, and that there is an efficient way of taking data from the active layer to a lower-cost inactive layer. The key to this is an meta-data architecture.
[-]
Ed Walsh
It seems a great opportunity for app vendors to optimize data placement since the app can intelligently anticipate needs
Jean-Luc Chatelain
@dvellante I think it is still very small but we are about do see an significant increase as witnessed/accelerated by the likes of SAP HANA
[+] Show Hidden Comments
[-]
Dave Vellante
So I agree w/that and I think it will explode. When you eliminate the mechanical bottleneck - great things will happen imo
[-]
Ed Walsh
Agreed seeing activity in Fast data space with SAP HANA and Pivotal GemfireXD
[-]
Dave Vellante
I think the development community has been dreaming of this day for a long time - a persistent memory extension that is cost-effective
[-]
Jean-Luc Chatelain
@dvellante I'll take pre-order for IME from you, we take Visa, MC & AMEX :-;
[-]
Ed Walsh
Interesting to see if app performance bottleneck moves from IO to CPU?
Jean-Luc Chatelain Indeed and that's an easier problem to solve but leads to some very interesting race conditions & locking contentions between threads. I love that this industry :-) !
Dave Vellante or the bottleneck moves to the network Ed
David Floyer
xflash (flash as an eXtension of memory) is likely to be the most cost-effective way of providing that kind of low latency.
[+] Show Hidden Comments
[-]
Dave Vellante
so describe in more detail - what exactly is x-flash?
[-]
David Floyer
xFlash is treating flash as an eXtension of DRAM - instead of going though the traditional SCSI stack multiple time to commit a write, Atomic writes do that only once and very quickly.
Dave Vellante so xflash by definition bypasses disk protocols?
[-]
David Floyer
Dave V - yes - no multi-phase disk protocols. The write can be completed in sub-microsecond time - down to 100 nanoseconds (not allowing for load on the system).
John Furrier
Whoever came up with xFlash is good. I love that name..
[+] Show Hidden Comments
[-]
David Floyer
Take a bow, Dave Vellante!
John Furrier how does xFlash handle xFiles ? couldn't resist
David Floyer with flash-backs?
Crowd Captain
How does virtualization play with the new hyperscale flash market in the enterprise
[+] Show Hidden Comments
[-]
Stuart Miniman
article posted today from @VMwareStorage discussing EMC ScaleIO with VMware http://cormachogan.com/2013/12/05/a-closer-look-at-emc-scaleio/
[-]
Ed Walsh
@virtualswede posted a good series of ScaleiO architecture and performance with virtualization http://purevirtual.eu/2013/11/29/microsoft-sql-server-testing-with-scaleio-on-vmware/
Stuart Miniman
We've seen huge growth in hybrid flash arrays, big launches of all flash arrays - can a solution be built of even lower latency which still having enterprise functionality?
[+] Show Hidden Comments
[-]
David Floyer
The business driver towards high performance systems will be applications that can radically lower the cost of doing business. Imagine you are OpenTable and you want to provide special offers to customers near a client restaurant. You have <1 sec!...
Crowd Captain
Don't forget to Vote for your favorite thread because the top thread will get the headline top position after the chat and the status of CrowdCaptain
[+] Show Hidden Comments
Jean-Luc Chatelain
In the end what ever technology paradigm accelerate the "time to results" for the "actionable insights" hungry enterprise will win and it must do in regardless of increase data volumes (and 3V etc...)
[+] Show Hidden Comments
[-]
David Floyer
Agreed - automation of actionable insights...
David Floyer
Good question Dave - for todays application designs - those who have very fast growing interconnected databases.
[+] Show Hidden Comments
[-]
Dave Vellante
traditional databases or new key value stores?
[-]
David Floyer
Both - traditional databases have great locking challenges and will need low latency quicker - KV stores will have greater transaction capability but have other constraints