Founded in March 2012
and heaqdquartered in Austin, TX - Mangstor designs PCIe SSD accelerators for
web-scale and large enterprise data centers to provide high-performance,
low-latency storage solutions that accelerate application workloads and data
storage operations running on traditional and virtualized servers. |
See also:-
Mangstor
- mentions on StorageSearch.com
2017 was a year like no
other in 40 years of SSD history.
If you're interested in Mangstor
you may also be interested in
Excelero,
PCIe SSDs and
rackmount SSDs. |
| .... |
| "... the emergence of
Excelero (in March 2017) made the early bright promises of low latency NVMe-like
SSD fabrics based on proprietary designs by Mangstor and EMC (DSSD) seem
expensive. You can now get similar performance with commodity hardware." |
| Mangstor raises
another $7 million - SSD news (August 2017) | | |
| .... |
Can a company with less employees than
cores in its SSD controller be significant for the market?
Yes -
here's why I rate Mangstor as a future Top SSD Company
by
Zsolt Kerekes.
editor - StorageSearch.com
- September 16, 2015
This month I had a long anticipated 90 minutes
conversation with Trevor Smith
CEO and co-founder of Mangstor .
Trevor
reminded me that we had spoken about the enterprise market over 20 years
earlier - when his company
Ross Technology
was marketing a competitive and faster multi-core alternative to Sun's 32 bit
SuperSPARC processor and when I was publishing the industry's buyer's guide
called the SPARC Product
Directory.
In the intervening period Trevor was VP Engineering at
ServerEngines
- which among other things in
2005 was one of
many companies working on to iSCSI and TCP/IP offload protocol engines and
which in
2010
was acquired by Emulex for its FCoE technology.
That background is
useful because if you want to understand the essential strength of Mangstor - it
helps to understand that when they decided to get into the enterprise SSD market
- and take advantage of the opportunities which would emerge from
standardization of software stacks around NVMe then 2 of the things which would
have deterred less ambitious startups:-
- writing the code for ultra low latency NVMe over GbE and InfiniBand
are
just some of the things they took in their stride without making too much of a
fuss about it.
Trevor told me he knew from his readings around the SSD
market (which included this site) that one of the keys to low jitter performane
was having lots of parallel operations aimed at the flash. In their initial
design they found a convenient 100 core 500MHz processor which was being used
for drone and military
applications - but the current processor design is their own IP.
Likewise
when they were looking for a way to run RDMA over NVMe fabrics - they thought
they would be able to use some standard code off the shelf. But as longtimers in
the enterprise connection market - they weren't deterred by writing their own.
It works best with Mellanox
adapters - which they chose because Trevor said they have the smallest buffer
sizes which guarantees the best latency.
How to quickly summarize
Mangstor?
Trevor said the company - which had a headcount of 41
employees when we spoke - and has received $15 million in 2 rounds of funding
- sees itself as being at the high end of the spectrum for NVMe PCIe SSD and
rackmount flash SSD performance.
There are some interesting comparisons
which can be made with previous companies which have dominated the lists of
fastest SSDs.
But
the key thing today is that most of the companies which used to have leading
edge products in this market space (fastest PCIe SSDs and fastest flash SSD
boxes) have been acquired
- and (for other reasons too) have lost their performance edginess - which
has left a space in the market for a newcomer like Mangstor - which can
introduce new product iterations at an impressive pace.
One of the
reasons is their
big controller
architecture - 100 core CPU architectural thinking. (Trevor said only about
70 are currently used.)
Where are they heading?
From a flash
perspective Trevor told me Mangstor is already on its 2nd memory supplier. The
switch was necessitated by market supply problems. But apparently Mangstor had
their design up and running with a new supplier - in just a few weeks.
Mangstor currently uses eMLC and a high rate of overprovisioning.
Having
played an active part in the enterprise processor and network commincations
markets for over 2 decades - Trevor has been pro-active in the areas of
compatibility testing and benchmarking.
Mangstor has even supplied
evaluation products to companies which - if I named them - are more likely to be
competitors than customers.
The reality is that no other company is
taking the same approach - and so no other company can match their performance
metrics.
It's likely that in a few years time - others will see the
advantages to be had from such a flexible no-compromize, end to end latency
oriented architecture.
But by then Mangstor will either have found a
use for those 30% of spare processor cores, or will be using more. Or maybe
turning up the clock speed and the number of lanes too.
On an NDA basis
I've seen some impressive competitive benchmarks - and even more impressive - a
low box to box latency which is delivered by their NVMe over GbE - which is
almost unbelievable and which will enable the viability of new applications
which until now have required risky new technologies like top of rack PCIe
fabrics - from PLX and
A3CUBE.
What's
the catch?
The world's fastest PCIe flash SSDs are pricey. And the
world's fastest NVMe AFA boxes are too.
For the right customers they
offer cleaner and easier to deploy solutions than any of the alternatives and
from a system perspective they can be cheaper too.
But like all new
startups - the reliability of Mangstor's new SSDs are unproven.
And
although their inherent low latency will enable simple
HA schemes to
be developed - without much loss of performance - these aren't part of the
proven solution set yet. That may take another few years to sort out. |
|
| In August 2015 -
Mangstor was awarded
Best of Show for "Most Innovative Flash Technology" during the
Flash Memory Summit. The Award
recognized Mangstor's
NX-Series
2U fast rackmount SSD which supports NVMe over Ethernet and InfiniBand
fabrics packaged through an RDMA cluster scale-out architecture. |
|
editor's comments:- June 2015 - Much has
changed re information from Mangstor in the 4 months since my earlier notes
about the company.
With a
datasheet of the
company's product now visible we can see the essential features of their MX6300
PCIe SSD cards are as follows:-
- form factor - FHHL
- type of flash - eMLC - capacity upto 2.7TB
- connectivity - PCIe Gen3 x8 (8GT/s)
- random R/W IOPS (70/30) - 700K IOPS (4KB)
- endurance - 7 DWPD
- R/W latency (4KB block) is 90 / 15 µS
- power consumption is 45W at 70/30 R/W
Unusual and noteworthy
characteristics of this product are:-
- Its controller includes a 100 core CPU.
- And - as predictable from an announcement in
August 2014 by
one of its memory suppliers -
Everspin - Mangstor's
SSD doesn't require supercaps for
power loss
protection as it uses ST-MRAM as its random on board cache.
Mangstor
also markets an InfiniBand
attached storage array - the NMX6320 - which can be used as a convenient box for
large scale deployments of its raw NVMe SSD technology.
In a press
release (May
2015) Mangstor provided more details.
The arrays are based on
Mangstor's Software Configurable MX6300-Series SSDs in combination with
Mangstor's TITAN storage stack. NVMe over RDMA read and write access to the
array provides servers with data at nearly identical latencies as if they were
accessing local PCIe SSDs. This allows compute clusters to realize the low
latency application benefits of PCIe SSDs without having to put an SSD in every
server. Typical configurations deliver millions of IOPS and throughput starting
at 14GB/S, with near linear scaling as arrays grow.
The idea of using
PCIe SSDs as the internal flash blocks inside fast SSD boxes isn't new. Earlier
pioneers in this technical approach were:-
NextIO,
Dolphin,
Fusion-io and
Virident (the
latter as a reference architecture rather than a product). But those earlier
boxes were either delivering a fast PCIe connected storage box (with a
proprietary software interface) or
legacy SAN FC
SAN or IB connectivity. Mangstor's products hope to leverage the software
standardization of NVMe to reach a new market which traditional styles of
fault tolerance
but coupled with a new type of SSDcentric software stack.
Many vendors
are working on this concept.
For an architecture perspective see
also:-
NVM
Express Over Fabrics (pdf) by
Intel (March 2015). |
|
who's who in SSD? - Mangstor
by
Zsolt Kerekes.
editor - StorageSearch.com
- February 2015
Although they had made an
announcement
about Everspin's
ST-MRAM in the lead up to the 2014
Flash Memory Summit - I
somehow missed that one. So Mangstor is a new - to me - SSD company.
Mangstor
contacted me recently because Trevor Smith,
Mangstor's CEO had been discussing the SSD market with some colleagues and
StorageSearch.com and my name came up. Is this the same
Zsolt Kerekes who
published the SPARC Product
Directory? Yes it is.
I had collaborated with Trevor Smith's
company ROSS
Technology in the early 1990s when ROSS was developing the market
for its independently designed 32 bit SPARC processor chip - called
hyperSPARC.
Trevor Smith was VP of Engineering at ROSS at the time.
So when I
first looked at the Mangstor's web site in early February 2015 - in response to
that "blast from the past" email it's not surprising I got the
impression that CPU cores, controllers and application performance seemed to be
at the heart of what the company was working on - but otherwise it still lacked
detail and had the air of a company still partly in stealth.
I'll be
learning more soon and will publish an update when Mangstor is ready to say
more.
I'd say it's worth keeping Mangstor on your watchlist if low
latency PCIe SSDs
are in your zone of interest. |
| . |
after AFAs -
what's next?
11 Key
Symmetries in SSD design
Exiting the
Astrological Age of Enterprise SSD Pricing
Decloaking
hidden and missing segments for enterprise flash
90% of
enterprise SSD companies have no good reasons to survive
|
| .. |
| .. |
|
| |
.. |
| The willingness of users to
entertain the idea of new architectures to solve big memory problems was shown
by the fact that an in-situ SSD processing pioneer - NGD Systems - moved up to
the #1 slot of companies which captivated the attention of StorageSearch readers
in Q4 2017 - replacing the long term AFA incumbent Pure Storage. |
| Top SSD Companies in Q4
2017 | | |
| .... |
 |
.. |
|
|
.. |
| what were
the big SSD ideas which emerged in 2016? |
"From Mangstor's
perspective, we are heavily invested in NVMe over Fabric (NVMf) technology for
which our NX-Series Storage Arrays are based.
With the ratification
of the NVMf specification this past June, flash devices such as SSDs and storage
arrays can now communicate over RDMA networks (such as RoCE or InfiniBand),
delivering the same high performance, low latency benefits as local attached
NVMe.
Storage arrays based on NVMf, such as our award-winning NX6320,
avoid the lower level SCSI transport layer, which results in faster data
throughput and accelerated data access making NVMf solutions ideal for
compute-intensive HPC and database applications."
Scott Harlin,
Director of MarComms - Mangstor
| | |
| .. |
|
|
| .. |
 |
| .. |
Mangstor gets $10 million
series B funding
for fastest NVMe flash SSDs |
Editor:- September 17, 2015 - Mangstor today
announced
it has closed $10 million in Series B funding which will be used to fuel
growth in sales and engineering and business development.
"This
investment confirms our leadership in NVMe over fabrics technology which
delivers an order of magnitude higher performance at the lowest latency verses
legacy iSCSI and Fiber Channel" said Trevor Smith
CEO and co-founder of Mangstor. | | |
.. |
|
|
.. |
|
|
.. |
|
here below - for comparison are 2 latency metrics from Mangstor plotted for
a single SSD with different queue depths |
 |
.. |
 |
at first glance - they look almost the same
and you think - so
what?
then you realize that the lower graph is NVMe over fabric
that
was my - Aha! moment |
|
.. |
|
|
.. |
|
| |
