what's the role for a Radian Memory SSD? would I be right in thinking
this...?
No. Turns out I was wrong. But I'm glad I asked.
by
Zsolt Kerekes,
editor - StorageSearch.com
- November 10, 2015 |
As long suffering readers know
I try to think of conceptual shortcuts to help place SSDs and related technical
concepts into mental buckets - sorted by simple design and architectural
features - which enable them to be placed in groups for later detailed
scrutiny.
Past examples of such technical filters include:-
So
a few months ago when I was thinking about Radian's RMS-250 (a
2.5" NVMe PCIe
SSD) it looked sufficiently out of place (compared to others in this form
factor) that I made a tentative decision to categorize it in a particular way.
As
this categorization would have far reaching effects (on my own future time
budget allocated to writing about the company, and the way that readers would
think about the product - if they adopted my cookie cutter) I thought I should
check my assumptions with the company first.
And I'm glad I did.
Because not only was my initial guess wrong - but other things which I gleaned
from this dialog (related to SSD and storage ancient history) enabled me to link
up some previously unconnected strands of market stories which I had
written about or read about decades before.
To save me time - and
to give you more detail - the rest of this article is cut and paste from
some emails between myself (that's the Zsolt part) and Mike Jadon,
CEO of Radian who recalled that we had spoken "about a dozen years
earlier (ouch!)" relative to his former company (Micro Memory) and the
PCI NVRAM line that they sold into storage OEMs.
Zsolt said - Would I be right in describing Radian's SSD as
the kind of product which a designer might think of having just one of? - as
an accelerator in an appliance rather than having large numbers of them in an
array?
2 things point in that direction
1 the power consumption
(25W) due to the memory place it at the upper end of the 2.5 NVMe SSD range.
And I know from talking to designers of arrays which have high SSD counts they
like the power controllable envelopes which are seen in array-friendly designs.
2
the multi-level software scalability. Does that scale effectively when the SSD
target is 4x, 8x and maybe 30x SSDs If I can get clarification on that it
will help me orient the appropriate positioning of the product for the right
kind of application considerations.
Mike answered each of these questions as follows.
re - Zsolt said - Would I be right in describing Radians SSD
as the kind of product which a designer might think of having just one of as an
accelerator in an appliance rather than having large numbers of them in an
array?
Mike said - Very much the opposite. Our team comes from Micro
Memory (supplier of devices into Storage OEMs), and from Storage array OEMs like
EMC/Isilon, Dot Hill, and Data Direct.
From the outset, the product
architecture was first designed to target arrays, and then we branched out to
targeting server-side applications.
Specifically, targeting arrays was an impetus for our 'offload engine'
architecture.
We are not a FTL at all, but any host based FTL tends
to consume lots of host resources and bottlenecks on host resources as devices
are added to a system.
Alternatively, treating the SSD as an offload
engine enables us to scale all of our perf. numbers (IOPS, bandwidth, latency,
QoS), but more importantly CPU/I-O utilization, as additional SSDs are added to
a system.
This concept has also fit well in server-side applications,
particularly hyperconverged, as it provides the maximum potential performance
but does so at the lowest utilization of host resources, making those host
resources available for resident applications.
re Zsolt's questions about power consumption?
Mike
said - We can also run the device lower than 25W, but our performance numbers
have been bench marked at max power utilization, hence we state 25W for
consistency.
Our bigger problem with supporting arrays today is that
most are still SAS/SATA based as the 2.5" NVMe ecosystem is still
developing. Many of the array companies, in particular the purpose-built AFA
companies, are moving to all NVMe arrays but the releases are more in the Q3 '16
time frame.
re Zsolt's question - the multi-level software scalability.
Does that scale effectively when the SSD target is 4x, 8x and maybe 30x SSDs?
Mike
said - Due to PCIe hardware platform limitations, we've only been able to
benchmark 7 SSDs on a single CPU?, and every metric scaled linearly. But in
theory our metrics should scale linearly to >20 SSDs.
Editor's footnotes.
The timing of these emails was around the
date of a SNIA conference presentation in which Radian talked on the theme
of Integrating
Cooperative Flash Management with SMR (Shingled Magnetic Recording) Technology
for Optimized Tiering in Hybrid Systems (pdf).
I didn't write about that presentation at the time - but the use of
Radian's SSD in a hybrid appliance is another thing which had led me to the
wrong assumption that it was a "one in a box" kind of product.
Mike
had some interesting comments on that (in the same emails) which touch on
Radian's views about the market direction. So I've pasted them here.
Mike
said - The announcement regarding SMR, the associated zbc standard, and its
alignment with our CFM (Cooperative Flash Management) technology is likely
bigger than it may first appear. SMR is going to take over cloud HDD storage,
and then make a significant impact into enterprise.
There's striking
similarities between SMR and NAND, and CFM exploits those similarities very well
so *systems* can now be optimized for both HDD and Flash.
Thinking about it from a different perspective, for the last five
years systems have been transitioning optimizations from HDD to Flash...
improving kernel schedulers, discarding legacy elevator algorithms, and the
introduction of the NVMe command set.
Continuing this evolution of
cleaning up inefficient legacy abstractions has to address the FTL - the
elephant in the room with regards abstractions, and one no longer necessary for
many host systems. Given the requirements for SMR, one of the main rationales
for FTLs, which is to emulate hard drives, is becoming an obsolete requirement.
Future systems can, for the first time, simultaneously be optimized
for both HDD and NAND media.
Editor's note - that's it. I hope you found it interesting. | | |
| ... |
SSD
news
the Top SSD
Companies
what's the
state of DWPD?
how fast can your SSD
run backwards?
what's
so very different about the design of Fusion-io's ioDrives / PCIe SSDs? |
| ... |
|
StorageSearch.com is published by
ACSL |
|
|
