introducing Memory Defined Softwareyes seriously - these words are
in the right order.
A new market for software which is strongly typed
to
new physical memory platforms and nvm-inside processors
while
unbound from the tyranny of memory virtualizable by storage.
by
Zsolt Kerekes,
editor - February 14, 2018 |
The existence of a market which provides
independent software support for solid state storage, SSDs and tiered memory
for enterprise use has a relatively short history (of only about 7 to 10 years)
compared to SSDs themselves. A tremendous amount has been accomplished in that
time (as you can see in the
SSD news
archives) as the computing industry transitioned from initially
shoe-horning SSDs into storage software models which had originally been written
for hard drives, then optimizing system software related code to detect and
bypass hardcoded rotating drive delay assumption workarounds which had been
buried in every type of application software and then finally creating a new
foundation of software primitives (NVMe) which began with the assumption that
storage could be solid state.
So far, so good - and there have been
some very talented companies which have revisited storage software assumptions
from inside the drive, outside in the array, in the interfaces, in the
associated stacks below, above, around and from every angle so that today you
can realistically expect to get operational characteristics from solid state
storage assets which are considerably better than whatever came before. And
although there is still work to be done the storage industry can congratulate
itself for collectively having done a good job despite at the start thrashing
around in a shambolic
state of disarray because the usual suspects didn't see the SSD avalanche
coming down the hill.
And it's because of the ubiquity of solid state
storage assets in the enterprise and the promise shown by early generations of
memory fabrics that the next phase of revolution in software is now underway -
which is how we get to memory defined software.
Let's backtrack
briefly to hardware - because hardware always comes first. Insofar (by way of
an apology in advance) that you can write all the software you like which
pretends that you're running a new computer business game on a new computing
platform - but you only start getting the benefits and the thrills by doing it
on the new hardware. Just over a year ago on this home page I wrote a blog -
after AFAs -
what's the next box? (cloud adapted memory systems) which hinted at the
kind of brew we should expect to see after the earlier pioneering percolators
of NVDIMM wars
had settled their territory disputes with alternative memories, tiered memory's
place relative to tiered storage and PCIe's transition from unruly invader to
settler in the territory of big memory fabrics which had upto not long before
been dominated by fast versions of very old interfaces (IB and GbE). (And just
to warn you that like previous land grabs - PCIe's position as a convenient
gateway into big memory spaces - is no more sacrosanct that what came before -
as Gen-Z may be a faster way to do things in future - although we have the
lessons of Infiniband versus Ethernet to show that sometimes the new does not
improve fast enough to displace what came before.)
You've had had
plenty of warning that something is coming.
What do I mean by
Memory Defined Software?
Simply this... Software which has been
deliberately written to take advantage of the computational realities of memory
with special characteristics in order to get behavior which was not possible
before. The special characteristics may take many forms:-
- nvm inside processors to enable instant reboot or context switches.
- trusted persistent memory which is used an as application dependent fast
look-up or code translation / computational acceleration / interpretive
resource
- memory which is bigger than traditional storage capacities - and which does
not break when you hit it with zillions of memory intensive operations which
require sub microsecond random read/modify/write/and move latencies.
- memory with embedded in-memory processing capability (achieved by FPGA or
ASIC).
At first some of the new memory defined software which is
designed to run on new memory systems may resemble the functional
characteristics of software which was developed to run on tiered memory systems
which include SSDs and flash as RAM virtualization. But just as storage software
evolved so that code written for flash environments could no longer run with
acceptable performance on HDD arrays - so too the split between true memory
defined software and software written for solid state storage installations
will become quickly apparent. I think sooner rather than later - as the stimulus
driving new memory code is coming from newer faster moving users who are more
nimble in their adoption of new platforms which solve data dependent problems
and doesn't carry the same decades long baggage and requirement for of
backwards compatibility.
This is the first time you've seen me use
the term "Memory Defined Software" in an article and it may feel not
quite right at first as it juggles in your brain space for a relationship
between SDS (Software Defined Storage) and the very similar sounding (but
entirely different) Software Defined Memory. But people have been experimenting
with software and architectures which are based around the Memory Defined
Software concept for a while to solve embedded problems. In the next stage of
the memoryfication of the enterprise I think it will become clearer that this is
a new big market opportunity so I thought it's time to recognize it for what
it is and give it its true name.
see also:-
are we ready
for infinitely faster RAM? | | |
.. |
|
|
.. |
|
|
.. |
 |
| |
.... |
 |
| .. |
|
|
| .. |
| Getting acquainted with the
needs of new big data apps |
Editor:- February 13, 2017 - The nature of
demands on storage and big memory systems has been changing.
A new
slideshare -
the
new storage applications by Nisha Talagala,
VP Engineering at Parallel
Machines provides a strategic overview of the raw characteristics of
dataflows which occur in new apps which involve advanced analytics,
machine learning and deep learning.
It describes how these new
trends differ to legacy enterprise storage patterns and discusses the
convergence of RDBMS and analytics towards continuous streams of enquiries.
And it shows why and where such new demands can only be satisfied by large
capacity persistent memory systems. |
 |
Among the many interesting observations:-
- Quality of service is different in the new apps.
Random access
is rare. Instead the data access patterns are heavily patterned and initiated by
operations in some sort of array or matrix.
- Correctness is hard to measure.
And determinism and repeatability
is not always present for streaming data. Because for example micro batch
processing can produce different results depending on arrival time versus event
time. (Computing the right answer too late is the wrong answer.) Nisha
concludes "Opportunities exist to significantly improve storage and memory
for these use cases by understanding and exploiting their priorities and
non-priorities for data." ...read
the article
SSD software news
where are we
heading with memory intensive systems? | | |
| .. |
|
|
| .. |
| GridGain
Systems - an example of Memory Defined Software |
A traditional example of Memory Defined Software
(and one of the easiest in my list of examples to understand) is the in-memory
computing software solution set from
GridGain Systems. In
its earliest implementations GridGain's products could be regarded as a
memory resident relational database. But by 2018 GridGain's
solutions had
expanded their ambitions and reach and could support multi node memory, cluster
snapshot, ACID transaction support, backup and persistence.
Moving
beyond traditional SQL in memory architectures GridGain has comfortably
expanded to interoperate with cloud based architectures and
IoT. | | |
| .. |
| Micron's
ACS - viewed as Memory Defined Software |
An outer limits example of Memory Defined
Software (but one which has profound promise for changing the boundaries of
compute platforms) is the work being done by
Micron to create
development platforms for its in-situ FGA inside memory array accelerators
ACS
datasheet (pdf).
The complexity of designing useful accelerators
for new applications using FPGAs which can leverage the dataflow benefits of
integrating the logic into offload intelligent memory is a large scale problem
which crosses many divides of competence. Micron has been working with machine
learning tool companies towards the holy grail of recompiling the nitty gritty
elements of big data problems into custom engines and architecture which can
be created in months rather than years of iterative design.
This
effort was outlined in a blog
Why
Memory Matters in Machine Learning for IoT (March 2018) - by Brad Spiers - Principal
Solutions Architect, Advanced Storage at Micron. | | |
.. |
|
|
.... |
| |
