|
|
|
where do ultrafast RAM SSDs and
companies like Kove fit in the market today? - part 2 of 2
by Zsolt Kerekes,
editor - StorageSearch.com
- September 2015
In the past year or so there have been many
developments in the SSD market which have improved application performance.
These include:-
- encroachments into the enterprise
DRAM socket by flash -
such as Diablo's
Memory1, by non flash nvms such as ReRAM - promised in
Intel's Optane
- shortening application latency in flash by various means in
SSD software (by 3rd
generation controllers and improved software architectures)
Which got
me wondering - where do ultrafast RAM SSD systems and companies like Kove fit
in the market today?
I put that question to John Overton,
CEO Kove in July 2015. Here are
some of the things he said.
"There are LOADS of quiet, NDA-based
applications for ultra-fast DRAM memory applications, that people may not spot
when thinking in terms of "storage" and Fibre Channel. Other
interconnects provide other options. Military and intelligence environments are
full of processing that *has* to be better not almost as good, where
lives depend upon reliability, speed, latency, determinacy."
"
In financial markets, as you mention, exchanges need determinacy. How about
persistent, HA memory mapped files? DRAM has profound improvement over SSDs any
time determinacy matters, because Flash needs software to protect itself, which
then introduces indeterminacy, and also as a function of concurrency. Databases
are well known application spaces."
"There are lots of
excellent strategies being introduced, as you mention on your site, including
Diablo as one illustration, that are getting to what Flash can do
physically/electronically. It's fine technology. But DRAM nearly eliminates
concurrency issues which Flash wasn't designed to do. If a bullet is flying
through the air, you'd better not have an arbitrary ~1 second outlier while the
software catches up. Embedding in silicon can help, but doesn't solve the
intrinsic jitter of Flash mechanics."
"One view (which you
rightly indicate) is legacy applications that have no other option. But,
similarly, those writing highly parallel applications benefit utterly and
brutally from the concurrency benefit that DRAM provides, under scale pressure
for read, write, both."
"Our markets (at Kove) don't find Flash attractive because it
doesn't deliver the economies of performance scaling for the applications we
enable. We maintain a storage interface for legacy purposes, for instance, but
it's one interface of many that we support, as we provide byte-addressability
that can build efficiently into block addressability. You can't do the reverse.
Similarly, for our applications, microseconds are not at issue but instead
nanoseconds." |
|
where do ultrafast RAM SSDs and
companies like Kove fit in the market today? - part 1 of 2
Editor:-
July 22, 2015 - it's been a few years since I last turned my attention to Kove.
The previous occasion - which I didn't write about on
StorageSearch.com - was when I was trying to puzzle out for myself what the
future technical directions might be for an SDS RAM SSD company - like Kove - in
an SSD ecosystem which had the possibilities of fast flash DIMMs - in particular
the
memory channel
SSD type - unveiled by Diablo
in 2013.
We now know that the first generation MCS SSDs were only a
little faster than the fastest
PCIe SSDs - so we'll
have to wait for the 2nd generation products (which were promised for 2015) to
see if the they really deliver a usable and scalable sub microsecond array
level latency.
My guess in 2013 was that while RAM SSDs could
simply scale up in capacity - the unreliability of worst case access times due
to cache misses in the MCS bridge chips would mean that the key benefit of RAM
SSDs - uncompromized and guaranteed
symmetric
latency couldn't be achieved - and therefore in Kove's key application areas -
this type of flash technology would not offer a worthwhile design change.
But
markets change. And recently a reader asked what I thought about ultrafast RAM
SSDs and companies like Kove in mid 2015?
Here's what I said.
Kove's
product served a narrowly defined but once common need in the enterprise
market...
To provide a low latency FC connected storage box which could
be shared by many servers in a business environment in which shaving 5 to 10
microseconds off the latency in a consistent fashion while maintaining simple
old fashioned legacy
apps and software – was a sound business objective.
Examples
being infrastructure in trading markets, exchanges etc.
Kove's solution was in effect an SDS configured with lots of RAM and
reliable server hardware and with software which emulated this function
reliably.
They weren't the first to offer such a RAM SSD approach and as we've
seen in the flash market – there are countless
software companies
providing similar functionality – but at higher latency.
This market space (for
ultra fast rackmount SSDs)
is changing in several ways.
1 – other ways of sharing remote
memory at low latency – such as
A3cube - whose provides
a low latency fabric across tens or hundreds of servers using a PCIe connected
shared memory with a low latency controller. With a small amount of software
adaptation it can encompass any legacy software app.
2 – arrays
of servers. The SDS market and SSD software market now enable the memory reach
of any single server to be larger – while using a mix of RAM,
hybrid DIMMs
and flash. These hyper converged systems will replace some of the legacy ways of
solving shared memory problems.
3 – memory channel SSDs. The
next generation of MCS from Diablo – with new software – will also
enable each server to have a much larger memory model with nearly DRAM latency
but flash like capacity. This doesn't solve the problem of sharing data between
servers. And also introduces the concept of longer response times if data is not
in the right cache. So some kind of fabric or SAN (FC or IB or 10GbE or PCIe
will still be needed.)
The role for a low latency FC connected SDS
box which looks like a RAM SSD – such as the Kove box will continue.
But users will require bigger capacities and there are other architectural
ways to solve similar problems. Products like A3cube's shared memory
connected via PCIe will become a bigger part of this solution space. But there
aren't enough systems deployed to enable safe modeling of the worst case
response times (yet). |
|
| . |
Who's
who in SSD? - by Zsolt Kerekes,
editor - November 28, 2011
Here's the short of it
Kove (a
top 20 SSD company)
markets the fastest rackmount SSD appliance available in a 4U form factor. If
you're researching products in this performance category you should also take a
look at
Texas Memory Systems,
Violin Memory and
Fusion-io.
Here's
the long version
Kove is 1 of 15 companies in the
fastest SSDs list,
and 1 of about 20 companies listed in the
RAM SSDs directory and
it's also in the top 20
SSD companies list. Kove is also listed in these guides too:-
FC SAN SSDs and
InfiniBand storage and
SSDs.
There's a growing realization in the enterprise market that
RAM SSDs will be a permanent part of the SSD toolkit in high transaction
volume datacenters which have any kind of traditional hierarchical data
architecture. And the only reason for buying them is to solve problems which
flash SSDs create or can't solve technically because of RAM's superior
latency, bandwidth, truly symmetric
IOPS,
better reliability (and sometimes lower floor price).
Kove's
Xpress Disk embodies how fast you can
go in an FC SAN /
IB connected RAM
storage appliance - with 8 microsends latency and nearly 30 gigabytes / s
bandwidth.
Who can afford this?
Most of you can't.
But if you're in the financial services industry and want an edge on
competitors delivering real-time market data - this is the type of product you
can't afford not to look at. Especially as it was recently (Oct 2011) shown
to be 12x faster than the previous fastest storage system when
performing a Market Snapshot
benchmark.
I currently
talk to more than 300 makers of SSDs and another 100 or so companies
which are closely enmeshed around the SSD ecosphere - which are all profiled
here on the mouse site.
I learn about new SSD companies every day,
including many in stealth mode. If you're interested in the growing
big picture of
the SSD market canvass - StorageSearch will help you along the way. Many
SSD company CEOs read our site too - and say they value our thought leading SSD
content - even when we say something that's not always comfortable to hear. I
hope you'll find it it useful too. |
| . |
|
|
In November 2010 -
Kove demonstrated a 4U
InfiniBand &
FC compatible
RAM SSD product line
called Xpress Disk - which can sustain
20GB/s throughput via 6x InfiniBand ports.
In June 2011 -
Kove announced that its
XPD2 4U 2TB RAM SSD had achieved the following performance:- 11.7 Million IOPS
in a single addressable space, 28.5GB/s bandwidth, along with round trip latency
of 6 microseconds for read and 8 microseconds for writes.
"The
next generation Kove Xpress Memory Disk is a continuation of our leadership in
high performance storage for those customers who need the absolute fastest
storage system available," states John Overton, Kove CEO. "The XPD2
connects directly to storage fabrics via standard Fibre Channel and InfiniBand
interconnects, while increasing performance density way beyond any other similar
storage alternatives."
In October 2011 -
Kove's
XPD2 was 12x faster than the
previous fastest system - in the Market Snapshot
benchmark audited by STAC.
In May 2012 - Kove published some
new
record latency numbers for its fast
RAM SSD - the XPD L2 -
which has achieved continuous and sustained 5 microsecond random storage read
and write when connected via 40Gb/s
InfiniBand adapters
from Mellanox . |
|
| |
 |
| .. |
| worst case response times
in DRAM arrays |
Do you know what the worst-case real-time
response of your electonic system is?
One of the interesting trends
in the computer market in the past 20 years is that although general purpose
enterprise servers have got better in terms of throughput - most of them are now
worse when it comes to latency.
It's easy to blame the processor
designers and the storage systems and those well known problems helped the
SSD accelerator market
grow to the level where things like
PCIe SSDs and
hybrid DIMMs have
become part of the standard server toolset. But what about the memory?
Server
memory based on DRAM isn't as good as it used to be. The details are documented
in a set of papers in my blog -
latency
loving reasons for fading out DRAM in the virtual memory slider mix. | | |
| .. |
 |
| ..... |
| sharpen your SSD R/W grid
latency to 5µS |
Editor:- May 9, 2012 - Kove has published some
new
record latency numbers for its fast
RAM SSD - the XPD L2 -
which has achieved continuous and sustained 5 microsecond random storage read
and write when connected via 40Gb/s
InfiniBand adapters
from Mellanox .
Kove's system has good
R/W symmetry
which the company says - is not subject to periodic performance jitter or "periodicity".
Even under constantly changing disk utilization, it delivers uniform,
predictable, and deterministic performance.
"The Kove XPD L2... allows high performance applications to use
storage as a weapon rather than accept it as a handicap," said Kove's
CEO, John
Overton. "We are pleased to set a new bar height for storage
latency." | | |
. |
 |
. |
|
|
. |
 |
| . |
|
|
. |
| "In February 2003
- 2 competing SSD companies:- Texas Memory Systems
and Imperial
Technology simultaneously announced the world's first terabyte class
SSD systems. The Tera-RamSan,
from TMS, provided 2 million IOPS, a 1024 gigabyte capacity, and 128 2-Gbit
Fibre Channel links. It required 2 racks and 5 kilowatts." |
| ...from:-
SSD Market
History | | | |
