a conversation with PLX about PCIe fabric, servers and SSDs StorageSearch.com
As a result they get a lot of top level visibility about what's happening in the enterprise computing market and when their company introduces something new - it can have a significant effect on many products.
Two months ago I had a conversation - about the SSD market from a PCIe interface angle - with Larry Chisvin, VP of strategic initiatives and David Hurd , Sr. Director, Corporate Communication at PLX Technology.
In this conversation I learned 2 new (to me) things which relate to the SSD market which I'm able to share with you now.
- a correction to one of my old SSD market penetration impact effect ideas
- and a new box - to join up the servers and SSDs in your life
a new PCIe fabric box - from PLX
PLX - whose primary business has been chips and related IP - is entering the systems business - with a PCIe fabric box
Anyone who has been reading articles by PLX knows that for the past 2-3 years they've been talking about the possibilities of using PCIe as a box to box fabric.
And I've talked to them about that subject in previous years too.
I said - this ExpressFabric solution idea which you talk about on your web site - I assume it's really just an outline capability of what someone could do with your chip interface and software technologies - if they choose to go in that direction them in that way...
It's not an actual product - is it?
It's just a concept diagram...
PLX said - well yes actually - that "box" you see in the diagram is a real product - and we've been sampling it under NDA to various customers.
Here's the exciting new thing I learned after probing. (Details may have changed by the time the product is launched.)
PLX started out with the intention of using the ExpressFabric box as a system design tool to help server oems and PCIe SSD designers get a feel for how they could expand their product offerings by scaling the fabric using PCIe instead of ethernet or InfiniBand.
And although the box - which has been sampling - doesn't use the next generation of silicon - which will be shipping later this year - and therefore isn't as fast as the production models will be - it provides developers with a workable platform which they can use to develop their software and test their architecture ideas.
What is the ExpressFabric box?
- 1U 32 port PCIe fabric - cabinet top box - to interconnect racks in a single cabinet
- internally there's a PCIe switch (3 switch chips each with 96 lanes) and a management server with software which controls the host to host communications - which supports line speed inter box RDMA
But the box has been designed to be production ready. And PLX told me that if customers decide to go to market shipping the PLX box as part of their solution - PLX has done the logistics planning to support that in a way that's competitive and viable.
Pricing will be in the region $5K to $10K (those details may change). The price will be set at a level which makes it competitive compared to using other fabric alternatives - which provide similar performance and architectures.
The great thing from the point of view of PCIe SSD companies - and users - is that the PLX box will provide a standard way to incrementally add more PCIe SSDs into a system - with high throughput, high availability and low latency - without having to resort to alien (non PCIe native) bridging technologies - which add delays and support complexities.
That's it - re the box anyway.
As I said before - details may change or I may have misread my notes.
But the outline story is that the "concept diagram" which has been at the heart of PLX presentations for the past several years - will sometime soon emerge as a box which you can use for linking servers and SSD resources together to create larger, faster and more reliable versions of the technology which many of you already like.
Related links:-
- PLX - profile page on StorageSearch.com
- PCIe SSDs directory on StorageSearch.com
- "ExpressFabric technology is initially targeted at small- to medium-sized cloud clusters with up to 1,000 nodes and 8 racks where the majority of the high-volume innovations are taking place." - ExpressFabric Wins Award (May 2014)
- "A3CUBE unveils PCIe memory fabric for 10,000 node-class architectures" - (February 2014)
In the first published version of this article (an SSD conversation with PLX) I placed this next section at the top of the page - before the new box. But then the next day I switched it around.
Most of you don't need to read it - as it's a marketing kind of thing.
On the other hand it just shows I can often be wrong too - in the summary ideas I carry around in my head. Even when I'm right about the lower level details. Due to not connecting the dots.
re SSD impact of unit shipments in the server market
It's common knowledge now that having SSD acceleration and the right kind of software in the enterprise impacts the nature and number of servers which are needed in the enterprise to perform workloads (in comparison to not having SSDs at all).
And hundreds of SSD companies and billions of dollars of revenue are now based on this concept.
Big cheer for the SSD companies and customers who made this happen.
But the market consequences of the idea - which I call SSD-CPU equivalence hadn't been analyzed (even by SSD companies who were enabling this trend) back in 2003 when I published one of my early SSD user value proposition based market sizing models - which looked at what would happen when the concept was more widely known.
The underlying technology concept - which was old even when I wrote about it - is that for a wide range of applications (as measured from an external reference point) - there is no little or no material performance difference between
- a single SSD accelerated CPU, or
- a single HDD based CPU (having the same architecture) but with a much higher clock speed (Nx) than the SSD system, or
- (Nx) more servers or cores of an HDD based CPU at the same clock speed as the reference SSD system
The background to the timing of my paper was that server CPU clock speeds had flatlined in the early years of the millenium due to limits imposed by architecture, the semiconductor and copper signal interface, and heat dissipation. This meant is was easier for CPU makers to add cores rather than add more GHz.
A long term consequence which I mentioned in my 2003 article - was that the world wouldn't need so many servers to do the same jobs.
And I said that would hit the server makers hard - which is why I said they would all have to become born again SSD marketers (when any one of them did) because there was no other future for them once the concept became widely known and easy to implement.
In those days - everything cost more - and the standard enterprise memory in SSDs was RAM - so the substitution effect mainly started in high end applications. And the early inroads made by SSDs in the enterprise were in customers who already owned the fastest servers - but who were still unhappy with the performance they were getting.
Anyway - what has that got to do with PLX and the market today? - you may ask.
I thought it would be a good opportunity for me to ask PLX - as they supply most of these server companies with the raw interface chips (regardless of the CPU type) - if they had seen any significant reductions in server volumes - which is what I expected from my original market model?
Zsolt - that's not exactly what we're seeing...
They didn't really want to go into too many details - about the total market for all server types - but were good enough to counter my headline suggestion - with their observation that instead of all server numbers shrinking down - they have instead been seeing high growth in a new type of server - called micro-servers.
That was a useful correction to my thinking about the server numbers - and made immediate sense to me - because this was another kind of user value proposition I had written about (later) in 2005 - where I had described the user value proposition for an ideal notebook PC. (If you're looking for that ideal SSD empowered notebook - despite the 9 years which have elapsed since I described what it should do - it hasn't appeared yet - and there's still no such product on the market.)
Anyway - the prime user value proposition was that you can run the apps on an SSD assisted notebook at desktop HDD speeds or faster - while using a slower clocked (and therefore lower power consumption) CPU.
The way the SSD notebook penetration model concepts apply to servers is this.
Instead of simply having dramatically less units of high end fast enterprise servers - the alternative (for users and cloud companies) is a mix and match approach of tactically using micro servers which have slower CPU clocks and which consumer much less electrical power.
So 1 SSD enhanced micro server can replace 1 non assisted high end server - is another alternative to - 1 SSD assisted high end server replacing 3 or more non SSD high end servers.
It depends on your workload, business and the cost of floor space - which option you'll choose. But the decline of fast clocked servers doesn't lead automatically to the decline of all servers. So in that respect I was wrong.
(And at some stage - the demand for servers will grow - because workloads don't stay the same from one decade to the next - and at some stage - new blue sky SSD enabled data industries will increase the demand for everything again.)
Anyway I found the market correction feedback from PLX useful - because I can run a long way on just a few simple new ideas each year.
Now many of you may be thinking - you already knew that (because unlike me you sometimes think about things which aren't SSDs).
In the quest to understand the SSD market (it's more like an Odyssey really - in which years of lost of time is spent visiting islands and meeting mysterious creatures which don't reappear in the story - but seem fascinatingly distracting at the time) - a lot of arcan articles get written.
Here are some related links:-