Data in motion #DellTechWorld

I (virtually) attended DTW this week and Michael Dell and others in their keynote segments mentioned that the new world involves both data at rest and data in motion. I was curious as to about this new concept of data in motion, so I spent some time looking into it.

AWS Lambda server less processing service and Apache Kafka probably best represents this idea of data in motion. Dell Boomi, IBM MQ, Google cloud Pub/Sub, etc. also provide similar services to Kafka.

With AWS Lambda, clients deposit data in object buckets and AWS automatically invokes some program, container, service, etc. to process that data and then the service goes away until the next data is deposited. Kafka is AWS Lambda on steroids.

Kafka is a completely open source (GitHub) system that’s run using a cluster of servers and provides a “message processing” system. A minimum Kafka cluster is 3 servers (containers, VMs or bare metal).

How does Kafka work

In Apache Kafka, you have producers, server/brokers and consumers. With Kafka, data comes in as events, with a key, values (essentially a bit stream, could be anything) and time-stamps which are created by producer clients and are automatically stored by Kafka servers or brokers and appended to topics (a sort of folder) in an ordered sequence. Topic events are then processed by consumer clients.

Topics are partitioned (sharded) using keys, and can be optionally replicated across a defined number of Kafka brokers within a cluster. Kafka clusters can span data centers , regions, clouds etc. Replication is done for fault tolerance. Topic partitioning provides scale out, distributed performance for Kafka.

Events can be simple messages for real time analysis or larger files for offline analysis. But they are all essentially produced, stored and consumed in an ordered, log like fashion.

Topic partitions can be multi-producer and multi-consumer. That is there can be 0, 1 or many producers of events in a topic (partition) and topic partitions can have 0, 1 or more consumers.

In Kafka, events are saved for a specified period and are not automatically jetisoned/deleted. As such, events can be read multiple times by consumers.

Kafka can also offer a guarantee that events are only processed once. Kafka can also guarantee that consumers of topic partitioned events always read events in arrival order.

Consumers register to see events they are interested in. As mentioned earlier, there can be multiple consumers of the same events. Consumers can take the form of micro services/containers, programs, systems, etc. When an event is stored, consumer clients registered for that event, get notified to process the event.

In Kafka producers and consumers are fully decoupled. They have no need to know about one another and indeed, can exist in different servers, clusters, data centers, etc. Event producers don’t wait on consumers. Event consumers are notified when an event is available and can do whatever processing is required for that event.

Kafka APIs

Kafka has APIs for:

  • Admin services API to provide monitoring and management of the Kafka cluster and services
  • Producers API to publish and create events
  • Consumers API to subscribe to read events
  • Kafka Streams API to supply higher level stream processing for events, such as micro services, with stateless processing, stateful processing, and within stateful processing, providing events within a (time) window. Events can be processed from one or more topics and used to transform (process) these into other events to be written to one or more topics. It supports per event processing with (2) millisecond latency for highly tuned systems. Streams use a Java API that can be deployed in containers, VMs, bare metal, in the cloud etc. Stream processing is not performed on the Kafka cluster but must be performed elsewhere. Kafka streams can be used to create advanced and complex data pipelines.
  • Kafka Connect API to supply the connections needed to get events from other outside, perhaps more traditional applications, environments, services into Kafka topics for processing and vice versa, output topic events to more traditional services. Connect services are available for many different applications, databases, systems, etc.. Connect can be used, for example, to provide a connection between an relational database and topics as well as connect topics to relational databases. You don’t code in Connect but rather provide declarative statements that define what data goes where.

Kafka is used in very many organizations (NY Times, LinkedIn, LINE, etc) to provide an almost, enterprise wide, all encompassing, processing bus where data comes in, is partitioned out to topics and then processed in real time or not. Kafka can be addictive. You start with a relatively small application and find uses for it throughout the company. Pretty soon, you are running your whole organization through Kafka.

Data in motion

So that’s an example of data in motion. Another way to think about Kafka and its data in motion is it’s represents the final step in the evolution of batch processing from mainframes of last century.

Batch processing of old, took a bunch of transactions, batched them together, and processed them one by one until the batch was done. With Kafka and similar systems, you essentially have a batch of one transaction and they provide all the framework and facilities needed to create, store and consume that single transaction (batch).

But in addition to this simplistic one transaction in, one process and one output. Kafka and other systems, provide a more general purpose system, with multiple transaction types (events) being created by multiple producers and being consumed by a multitude of processes, that can each produce one or more outputs which could be other events to start the process over again. This create event, process, create event, process, could go on ad infinitum.

And that’s what a data pipeline looks like. Event data comes in, it’s processed (filtered, aggregated, merged, etc.) and generates a different event which causes more processing, which creates other events, which causes other processing…..

And that’s data in motion.

Photo Credit(s):

All graphics and photos are from Apache Kafka website

An Open Source Powered Leg

I read an article a couple of weeks back about an Open Source Bionic Leg, which was reporting on research began as a NSF funded project at the University of Michigan (UoM), with collaboration from Northwestern, University of Texas at Dallas and CMU. UofM has a website that provides everything you need to build your own open source leg (OSL) leg at

The challenge in human prosthetics these days is that all research is done in silos. Much of it is proprietary and only available within corporations but even university research has been hampered by the lack of a standard platform that could be used to develop new components and ideas on.

The real difficulty is defining the control logic (code). The OSL project is intended to resolve this lack of a platform by providing everything a researcher (hobbyist, or amputee) needs to build their own, at home or in the lab.

The website includes a parts lists and STEP files as well as an estimated cost ($28.5K) to build your own powered prosthetic leg. They also have a Excel spread sheet with all the parts listed, including part numbers and links to where they can be ordered (McMaster-Carr, SolidWorks, & Dephy)

They also show how to build a leg with a short youtube video of how to assemble the whole leg as well as details for each subassembly with separate how-to videos for each.

The open source leg makes use of code from FlexSEA (Flexible Scaleable Electronics Architecture) and Dephy. FlexSea was originally developed by Jean-Francois (JF) Duval while he was at MIT for his doctoral thesis. He has since joined Dephy a robotics design firm. The open source leg project uses FlexSea/Dephy code for its servo control mechanisms.

There is a GitHub Python, MatLab and C control library repo with all the code. The open source leg website also includes instructions, scripts and an image file which can be used to build your own RaspberryPi (4) controller for the leg.

The two (ankle and knee) servos are USB connected to the RPi. There are also other sensors such as the joint (servo-motor) encoders and a six axis load sensor I2C connected to the RPi. Each servo has its own 950mAh battery.

On the OSL website’s control page one can see these servos in action (with short youtube segments). They also provide instructions on how to use the open source control library to take the servo mechanisms through their paces.

Although on the OSL website’s control page I didn’t see anything which put the whole leg together to make use of it in a real world application. They did show on the Data page a youtube video with the OSL attached to a person and being used to walk up and down stairs, inclines and walking across a floor.


Seeing as how the OSL website included STEP and PDF files for all the (machined) parts which represent $15.6K of the $28.5K, if one really wanted to do this on the cheap, one could just 3D print these parts in plastic. It would obviously not suffice mechanically for real use, but it could provide a platform for testing and developing control logic. At some point one could upgrade some or all of the plastic 3D printed parts to something more durable for use in human trials.

Another option is to purchase multiple sets of parts. The OSL website also showed price estimates for purchasing two sets of ankle and knee parts. But I’d imagine if one was so inclined, a number of researchers (hobbyists or amputees) could get together and order multiple sets of parts for reduced prices.

It’s also possible, with a lot of work, that the open source leg could be redesigned to support an open source arm-hand mechanism. This is where having 3D printed plastic parts could be extremely useful in helping to redesign the leg into an arm-hand.

Photo Credit(s):

Where should IoT data be processed – part 2

I wrote a post a while back on Where IOT data should be processed – part 1. We will get back to that post in a moment, but recently I read an article (How big data forced the hunt for ET intelligence to evolve) that mentioned after 20 years, they were shutting down SETI@home.

SETI@home was a crowdsourced computational network that took snippets of radio spectrum, sent them to 1000s of home computers to be analyzed during idle computer time, once processed the analysis was sent back to SETI@home. It was one of the first to use a crowdsourced approach to perform data processing. The data was collected at a radio telescope, sent to SETI@home and distributed from there.

6 Factors for IOT data processing

In my post I talked about 6 factors that should help determine where data is processed. Those 6 factors included

  • Data size which is a measure of the amount (GB, TB or PBs) of data that is being generated at an IOT node
  • Data pipe availability, which is all about the networking bandwidth that’s available at the IOT node. If we are talking some sort of low-bandwidth networking access then it probably makes sense to process the data more locally and send only results of processing up the stack.
  • Processing criticality which indicates how important is the processing of the data. If the processing could save a life then maybe it should be done as close as possible to where the data is generated. If the data processing is less critical it could perhaps be done at other nodes in an IOT network
  • Processing time and infrastructure cost which is all about what sort of computational resources are required to perform the processing and how much would it cost. If processing of the data is to undergo multiple passes or requires multi-core CPUs or GPUs, moving data off the IoT node and onto a more comprehensive server to process it, could make sense.
  • Compliance, governance and archive requirements, which discussed the potential need for all data to be available for regulatory audits and as such may need to be available at a central location anyway so why not perform processing there.
  • Data information funnel, which talked about the fact that an IoT network should be configured in layers and that each layer in the stack should probably be responsible for some portion of the data processing needed by the overall system, if nothing more than compressing the information before it is sent elsewhere.

Now that I review the list, the last, Data information funnel, factor really should be a function of the other factors rather than a separate factor.

In that blog post I promised to follow it up with some examples of the logic applied to real world problems. SETI is the first one I’ve seen in the literature

SETI’s IoT processing problem

Closeup front view of one antenna of the Allan Telescope Array, a radio telescope for combined radio astronomy and SETI (Search for Extraterrestrial Intelligence) research being built by the University of California at Berkeley, outside San Francisco. The first phase, consisting of 42 6 meter dish antennas like the one shown here, was completed in 2007. Eventually it will have 350 antennas. This type of antenna is called an offset Gregorian design. The incoming radio waves are reflected by the large parabolic dish onto a secondary concave parabolic reflector in front of the dish, and then into a feed horn. A metal shroud can be seen along the bottom of the secondary reflector which shields the antenna from ground noise. It covers the frequency range from 0.5 to 11.2 GHz.

The SETI researchers found that “The telescopes are now capable of producing so much data that it’s not possible to get that volume of data out to volunteers,” And “The discovery space is in these massive, massive data streams. And it’s just not efficient to distribute many terabits per second out to volunteers all over the world. It’s more efficient for that data processing to happen at the actual observatory.”

So they moved the data processing for the SETI IoT network from being distributed out to home computers throughout the world to being done at the (telescope) source where the data was originally generated.

This decision seems to rely on a couple of the factors above. Namely the pipe availability and data size factors. They had to move processing because no pipes existed to send Tb of data to 1000s of home computers. And finally, the processing time and infrastructure cost has come down so much, that it was just easier to do the processing onsite.

It doesn’t seem like processing criticality or compliance-governance-archive had any bearing on the decision.

So there’s the first example that seems to fit well into our data processing framework.


We ought to be able to come up with a formula that uses all these factors and comes up to with a yes or no as to whether to process the data on the node or not.

Photo Credit(s)

Undersea datacenter in our future?

Read about Microsoft’s Project Natick Phase 2 this past week. Microsoft submerged a steel encased tube filled with servers, storage and compute for 2 years in the UK and just took it out of the water this past July. We’ve written before about underwater and in space data centers (see our IT in space post)

Project Natick’s Phase 2 underwater data center had 12 racks with 864 servers and 27.5PB of disk storage and was connected to the nearby Orkney island’s power grid (250Kw) and networking infrastructure. The Orkney’s islands are located off the NE coast of the Scotland and its power grid is 100% renewable, using tidal, solar and wind power. During the data center test, Orkney was able was able to power the data center, the islands and still provide power back to the Scottish power grid.

More reliable underwater

According to early reports, the servers in the underwater data center had 1/8th the failures that a control data center, on land, had. Microsoft attributes the enhanced server reliability to the use of a 100% Nitrogen (at 1 atmosphere pressure) rather than normal air and the lack of any humans to jostle the equipment/disturb the environment.

It’s also likely that the temperature variability present in a normal, on the surface of the earth, data center was measurably less than for a data center on the sea floor. If this were true, that could also help explain its better reliability.

Why underwater?

It’s all about cooling modern servers (and storage). According to NREL ( USA National Renewable Energy Lab), most data centers operate at 1.8 PUE (power use efficiency) that is, using 180% of the power required for the servers, storage and networking equipment. The other 80% is used mainly for cooling electronics, but also includes lighting, HVAC, and other essential services for humans. NREL says that high efficiency data centers can achieve a PUE of 1.2.

PUE for Project Natick Phase 2 data center was reported to be 1.07. The only additional electricity needed would probably be power for cooling.

Cooling for the Project Natick Phase 2 data center used seawater pumped through the back of server racks. The data center was placed on the seafloor at 35m (117ft) deep.

It kind looked like a submarine. According to Microsoft, the data center was contracted for, built and deployed in under 90 days. The intent was to have the data center be smaller than a standard ISO shipping container. The data center was driven ontop of an 18 wheeler, from where it was built to the Orkney Island, including ferry crossings. It was placed on a triangular support, towed out to see and deposited on the seafloor.

While 864 servers and 27.5PB of storage seem like a lot to most of us, for Microsoft Azure it’s too small to be used as a regional zone. But for (large) edge deployments. something this size or (10X) smaller might be just the thing.

Microsoft notes that 1/2 the world’s population lives within 200km (120mi) of the ocean. So there’s a ready supply of people and businesses that could take advantage of any underwater data center.

And of course, such a structure when laid on the bottom of the ocean floor, could create an artificial reef (if left in place long enough). Artificial reefs have been made out of ocean oil rigs, sunken war ships and large chunks of steel/concrete. So a underwater data center could do so just as well. And maybe the heating coming from the data center cooling pumps would foster even more coral life.

Microsoft plans Project Natick Phase 3 to be a full Azure AZ that will be deployed underwater which will include about 12 Phase 2 datacenter pressurized units.

Photo Credits:

New PCM could supply 36PB of memory to CPUs

Read an article this past week on how quantum geometry can enable a new form of PCM (phase change memory) that is based on stacks of metallic layers (SciTech Daily article: Berry curvature memory: quantum geometry enables information storage in metallic layers), That article referred to a Nature article (Berry curvature memory through electrically driven stacking transitions) behind a paywall but I found a pre-print of it, Berry curvature memory through electrically driven stacking transitions.

Figure 1| Signatures of two different electrically-driven phase transitions in WTe2. a, Side view (b–c plane) of unit cell showing possible stacking orders in WTe2 (monoclinic 1T’, polar orthorhombic Td,↑ or Td,↓) and schematics of their Berry curvature distributions in momentum space. The spontaneous polarization and the Berry curvature dipole are labelled as P and D, respectively. The yellow spheres refer to W atoms while the black spheres represent Te atoms. b, Schematic of dual-gate h-BN capped WTe2 evice. c, Electrical conductance G with rectangular-shape hysteresis (labeled as Type I) induced by external doping at 80 K. Pure doping was applied following Vt/dt = Vb/db under a scan sequence indicated by black arrows. d, Electrical conductance G with butterfly-shape switching (labeled as Type II) driven by electric field at 80 K. Pure E field was applied following -Vt/dt = Vb/db under a scan sequence indicated by black arrows. Positive E⊥ is defined along +c axis. Based on the distinct hysteresis observations in c and d, two different phase transitions can be induced by different gating configurations.

The number one challenge in IT today,is that data just keeps growing. 2+ Exabytes today and much more tomorrow.

All that information takes storage, bandwidth and ultimately some form of computation to take advantage of it. While computation, bandwidth, and storage density all keep going up, at some point the energy required to read, write, transmit and compute over all these Exabytes of data will become a significant burden to the world.

PCM and other forms of NVM such as Intel’s Optane PMEM, have brought a step change in how much data can be stored close to server CPUs today. And as, Optane PMEM doesn’t require refresh, it has also reduced the energy required to store and sustain that data over DRAM. I have no doubt that density, energy consumption and performance will continue to improve for these devices over the coming years, if not decades.

In the mean time, researchers are actively pursuing different classes of material that could replace or improve on PCM with even less power, better performance and higher densities. Berry Curvature Memory is the first I’ve seen that has several significant advantages over PCM today.

Berry Curvature Memory (BCM)

I spent some time trying to gain an understanding of Berry Curvatures.. As much as I can gather it’s a quantum-mechanical geometric effect that quantifies the topological characteristics of the entanglement of electrons in a crystal. Suffice it to say, it’s something that can be measured as a elecro-magnetic field that provides phase transitions (on-off) in a metallic crystal at the topological level. 

In the case of BCM, they used three to five atomically thin, mono-layers of  WTe2 (Tungsten Ditelluride), a Type II  Weyl semi-metal that exhibits super conductivity, high magneto-resistance, and the ability to alter interlayer sliding through the use of terahertz (Thz) radiation. 

It appears that by using BCM in a memory, 

Fig. 4| Layer-parity selective Berry curvature memory behavior in Td,↑ to Td,↓ stacking transition. a,
The nonlinear Hall effect measurement schematics. An applied current flow along the a axis results in the generation of nonlinear Hall voltage along the b axis, proportional to the Berry curvature dipole strength at the Fermi level. b, Quadratic amplitude of nonlinear transverse voltage at 2ω as a function of longitudinal current at ω. c, d, Electric field dependent longitudinal conductance (upper figure) and nonlinear Hall signal (lower figure) in trilayer WTe2 and four-layer WTe2 respectively. Though similar butterfly-shape hysteresis in longitudinal conductance are observed, the sign of the nonlinear Hall signal was observed to be reversed in the trilayer while maintaining unchanged in the four-layer crystal. Because the nonlinear Hall signal (V⊥,2ω / (V//,ω)2 ) is proportional to Berry curvature dipole strength, it indicates the flipping of Berry curvature dipole only occurs in trilayer. e, Schematics of layer-parity selective symmetry operations effectively transforming Td,↑ to Td,↓. The interlayer sliding transition between these two ferroelectric stackings is equivalent to an inversion operation in odd layer while a mirror operation respect to the ab plane in even layer. f, g, Calculated Berry curvature Ωc distribution in 2D Brillouin zone at the Fermi level for Td,↑ and Td,↓ in trilayer and four-layer WTe2. The symmetry operation analysis and first principle calculations confirm Berry curvature and its dipole sign reversal in trilayer while invariant in four-layer, leading to the observed layer-parity selective nonlinear Hall memory behavior.
  • To alter a memory cell takes “a few meV/unit cell, two orders of magnitude less than conventional bond rearrangement in phase change materials” (PCM). Which in laymen’s terms says it takes 100X less energy to change a bit than PCM.
  • To alter a memory cell it uses terahertz radiation (Thz) this uses pulses of light or other electromagnetic radiation whose wavelength is on the order of picoseconds or less to change a memory cell. This is 1000X faster than other PCM that exist today.
  • To construct a BCM memory cell takes between 13 and 16  atoms of W and Te2 constructed of 3 to 5 layers of atomically thin, WTe2 semi-metal.

While it’s hard to see in the figure above, the way this memory works is that the inner layer slides left to right with respect to the picture and it’s this realignment of atoms between the three or five layers that give rise to the changes in the Berry Curvature phase space or provide on-off switching.

To get from the lab to product is a long road but the fact that it has density, energy and speed advantages measured in multiple orders of magnitude certainly bode well for it’s potential to disrupt current PCM technologies.

Potential problems with BCM

Nonetheless, even though it exhibits superior performance characteritics with respect to PCM, there are a number of possible issues that could limit it’s use.

One concern (on my part) is that the inner-layer sliding may induce some sort of fatigue. Although, I’ve heard that mechanical fatigue at the atomic level is not nearly as much of a concern as one sees in (> atomic scale and) larger structures. I must assume this would induce some stress and as such, limit the (Write cycles) endurance of BCM.

Another possible concern is how to shrink size of the Thz radiation required to only write a small area of the material. Yes one memory cell can be measured bi the width of 3 atoms, but the next question is how far away do I need to place the next memory cell. The laser used in BCM focused down to ~1.5 μm. At this size it’s 1,000X bigger than the BCM memory cell width (~1.5 nm).

Yet another potential problem is that current BCM must be embedded in a continuous flow of liquid nitrogen (@80K). Unclear how much of a requirement this temperature is for BCM to function. But there are no computers nowadays that require this level of cooling.

Figure 3| Td,↑ to Td,↓ stacking transitions with preserved crystal orientation in Type II hysteresis. a,
in-situ SHG intensity evolution in Type II phase transition, driven by a pure E field sweep on a four-layer and a five-layer Td-WTe2 devices (indicated by the arrows). Both show butterfly-shape SHG intensity hysteresis responses as a signature of ferroelectric switching between upward and downward polarization phases. The intensity minima at turning points in four-layer and five-layer crystals show significant difference in magnitude, consistent with the layer dependent SHG contrast in 1T’ stacking. This suggests changes in stacking structures take place during the Type II phase transition, which may involve 1T’ stacking as the intermediate state. b, Raman spectra of both interlayer and intralayer vibrations of fully poled upward and downward polarization phases in the 5L sample, showing nearly identical characteristic phonons of polar Td crystals. c, SHG intensity of fully poled upward and downward polarization phases as a function of analyzer polarization angle, with fixed incident polarization along p direction (or b axis). Both the polarization patterns and lobe orientations of these two phases are almost the same and can be well fitted based on the second order susceptibility matrix of Pm space group (Supplementary Information Section I). These observations reveal the transition between Td,↑ and Td,↓ stacking orders is the origin of
Type II phase transition, through which the crystal orientations are preserved.

Finally, from my perspective, can such a memory can be stacked vertically, with a higher number of layers. Yes there are three to five layers of the WTe2 used in BCM but can you put another three to five layers on top of that, and then another. Although the researchers used three, four and five layer configurations, it appears that although it changed the amplitude of the Berry Curvature effect, it didn’t seem to add more states to the transition.. If we were to more layers of WTe2 would we be able to discern say 16 different states (like QLC NAND today).


So there’s a ways to go to productize BCM. But, aside from eliminating the low-temperature requirements, everything else looks pretty doable, at least to me.

I think it would open up a whole new dimension of applications, if we had say 60TB of memory to compute with, don’t you think?


[Updated the title from 60TB to PB to 36PB as I understood how much memory PMEM can provide today…, the Eds.]

Photo Credit(s):

Storage that provides 100% performance at 99% full

A couple of weeks back we were talking with Qumulo at Storage Field Day 20 (SFD20) and they made mention that they were able to provide 100% performance at 99% full. Please see their video session during SFD20 (which can be seen here). I was a bit incredulous of this seeing as how every other modern storage system performance degrades long before they get to 99% capacity.

So I asked them to explain how this was possible. But before we get to that a little background on modern storage systems would be warranted.

The perils of log structured file systems

Most modern storage systems use a log structured file system where when they write data they write it to a sequential log and use a virtual addressing scheme to show where the data is located for that address, creating a (data) log of written blocks.

However, when data is overwritten, it leaves gaps in these data logs. These gaps need to be somehow recycled (squeezed out) in order to be able to be consumed as storage capacity. This recycling process is commonly called “garbage collection”.

Garbage collection does its work by reading heavily gapped log files and re-writing the old, but still current, data into a new log. This frees up those gaps to be reused. But garbage collection like this takes reading and writing of logs to free up space.

Now as log structured file systems get (70-80-90%) full, they need to spend more and more system time and effort (=performance) garbage collecting . This takes system (IO) performance away from normal host IO activity. Which is why I didn’t believe that Qumulo could offer 100% IO performance at 99% full.

But there was always another way to supply storage virtualization (read snapshotting) besides log files. Yes it might involve more metadata (table) management, but what it takes in more metadata, it gives back by requiring no garbage collection.

How Qumulo does without garbage collection

Qumulo has a scaled block store for a back end of their file and object cluster store. And yes it’s still a virtualized block store BUT it’s not a log structured file store.

It seems that there’s a virtual-to-physical mapping table that is used by Qumulo to determine the physical address of any virtual block in the file system. And files are allocated to virtual blocks directly through the use of B-tree metadata. These B-trees indicate which virtual blocks are in use by a file and its snapshots

If a host overwrites a data block. The block can be freed (if not being used in a snapshot) and placed on a freed block list and a new block is allocated in its place. The file’s allocated blocks b-tree is updated to reflect the new block and that’s it.

For snapshots, Qumulo uses something they call “write-out-of-place” process when data that a snapshot points to is overwritten. Again, it appears as if snapshots are some extra metadata associated with a file’s B-tree that defines the data in the snapshot.

The problem comes in when a file is deleted. If it’s a big enough file (TB-PB?), there could be millions to billions of blocks that have to be freed up. This would take entirely too long for a delete command, so this is done in the background. Qumulo calls this “reclaim delete“. So a delete of a big file unlinks the block B-tree from the directory and puts it on this reclaim delete work queue to free up these blocks later. Similarly, when a big snapshot is deleted, Qumulo performs a background process called “reclaim snapshot” for snapshot unique blocks.

As can be seen (it’s very hard to see given the coloration of the chart) from this screen shot of Qumulo’s session at SFD20, reclaim delete and reclaim snapshot are being done concurrently (in the background) with normal system IO. What’s interesting to note here is that reclaim IO (delete and snapshots) are going on all the time during the customers actual work. Why the write throughput drops significantly doing the the 27-29 of July is hard to understand. But the one case where it’s most serious (middle of July 28) reclaim IO also drops significantly. If reclaim IO were impacting write performance I would have expected it to have gone higher when write throughput went lower. But that’s not the case. From what I can see in the above reclaim IO has no impact on read or write throughput at this customer.

So essentially, by using a backing block store that does no garbage collection (not using a log structured file system), Qumulo is able to offer 100% system IO performance at 99% full – woah.

AI ML DL hardware performance results from MLPerf

Read an article a couple of weeks back from IEEE Spectrum, New Records for AI Training which discussed recent MLPerf v0.7 performance results. The article mentioned that MLPerf performance on its benchmarks has increased by ~2.7X in the last year alone.

The MLPerf organization was started back in 2018 to supply machine learning workload performance results, somewhat like what SPEC and TPC did for NFS and transaction processing. The MLPerf organization documented their philosophy in a paper

As far as I can tell, MLPerf is the only benchmark currently available to show hardware system performance on AI training and inferencing. Below we report on MLPerf training results.

MLPerf also reports on both closed and open division benchmark results. Closed division submission all use the same software algorithms for each workload submission. This way one can compare workload performance across different hardware systems. Open division results can make use of any algorithm to achieve the desired results on the problem set. We report on MLPerf closed division results below.

Current MLPerf v0.7 (open and closed division) training results are available online (on GitHub) and are summarized in a training results page on their web site.

MLPerf v0.7 workload changes

The MLPerf team added a few new workloads and upped the game of another benchmark for V0.7

  • Recommendation DLRM: a replacement for what was used in MLPerf v0.6 and is from Facebook providing more parallelism in training for recommendations.
  • Wikipedia BERT: an addition to what was used in MLPerf v0.6 and is a new natural language processing (N?P) frontend, trained on Wikipedia which is used with other language processing capabilities.
  • Go MiniGo: an enhancement to MLPerf v0.6 MiniGo accuracy requirements and uses reinforcement learning to learn to play Go well enough to achieve a 50% win rate. For v0.7, they now use a full sized, 19X19 Go board and upped the win rate requirement to 50%.

MiniGo Results

A couple of items of note for the MiniGo results. There are essentially 3 different architectures represented in the above: NVIDIA DGX series (DGX A100, DGX-2H, DGX-1), Google TPUs (V4 and V3) and Intel (8 server nodes with Copper Lake-6 CPUs).

Google TPUs are considered internal and are only available to Google, its hardware partners or on GCP. Although MLPerf include GCP TPU system results for other workloads, there were none submitted for MiniGo.

The Intel system is a preview of their latest gen Copper Lake chips, which may not be commercially available yet. On the other hand, all NVIDIA systems are commercially available and can be deployed in your data center today.

As one can see in the above, NVIDIA systems swept the first 3 positions on our Top 10 MiniGo chart. A DGX A100 came in at #1, reaching a 50% win rate at MiniGo in mere 17 seconds using 448 CPUs and 1792 A100 GPUs. Coming in at #2 at 30 seconds was another DGX A100 using 64 CPUs and 256 A100 GPUs. And at #3 at 35 seconds was a DGX-2H using 64 CPUs and 512 V100 GPUs.

Next at #4 at 151 seconds was a Google TPU system with 64 TPUv4 accelerators (unclear how many CPUs, if any are used, results show 0). Note, an 8-node Intel server with the 32 CPUs (4/node) using the latest gen Copper Lake (-6) CPU came in at #7 using 409 seconds to achieve the training results.

There are 6 other MLPerf workloads including DLRM and BERT mentioned above. Each of these deserve their own discussion on top ten results. Alas, they will need to wait for another time and I will cover all of them in future posts.


Nowadays, with much of IT turning to AI ML DL to provide critical services, it’s more important than ever to understand what can and can’t be done with available hardware. The fact that one can train a model to play decent Go in 17 seconds on a large DGX A100 cluster and under 7 minutes on an 8-node, leading edge, Intel server cluster is pretty impressive.

Despite MLPerf’s best efforts, it’s still tough to compare ML performance across systems when there’s so much diversity in the underlying hardware, especially in GPU, TPU and CPU counts. IMHO, it would be very useful to have a single GPU , TPU or CPU system submission requirement for each workload. That way one could compare how well each hardware element can perform the workload in isolation.

Nonetheless, the MLPerf suite of benchmarks provides a great first step in understanding what today’s hardware can accomplish in ML training (and inferencing).


Photo Credits:

Can we back up a PB?

Tradition says no way. IT backup history says not on your life. Common sense would say never in a million years.

Most organizations with PB of data or more, depend on remote replication to protect against data center outage or massive loss of data. This of course costs ~2X your original data center. And for some organizations one copy is not enough, so ~3X .

I don’t know what a PB scale data storage costs these days but I can’t believe it’s under a couple Million $ USD in hw and sw costs and probably at least another Million or so in OpEx/year. Multiply that by 2 or 3X and you’re now talking real money.

How could backup help?

Well for one you wouldn’t need replicas, so that would cut your hw & sw acquisition costs by a factor of 2 or 3. But backup storage is not free either. So you’d probably need to add back 30-50% of the original data center in hw & sw costs for backups.

You certainly wouldn’t need as many admins. And power for backup storage should also be substantially less. So maybe your OpEx would only be 1.5X in total for the original PB and its backups.

But what could possibly back up a PB of data?

We were talking with Igneous at Cloud Field Day 8 (CFD8, see their video here)  a couple of weeks back and they said they could and do backup PBs of data for customers today. A while back, e also talked with them on a GreyBeards on Storage podcast.

The problems with backing up a PB seem insurmountable. First you have to be able to scan a PB of data. This means looking into multiple file systems on many different hardware platforms, across potentially multiple data centers, and that’s just to get a baseline of what all needs to be backed up.

Then at some point you actually have to store all that data on backup storage. So, to gain some cost advantage, you’d want to compress and deduplicate a PB of data, so that the first full backup wouldn’t take a full PB of backup storage.

Then of course you have to transfer a PB of data to your backup storage, in something that wouldn’t take months to perform. And that just gets you the first full backup.

Next, comes the daily scan of what’s changed. This has to re-scan your PB of data to find that 100TB or so, that’s changed over the last 24 hrs. Sometime after that scan completes, then all that 100TB or so of changed data needs to be compressed, deduped and transferred again to backup storage

And if that’s not enough, you have to do it all over again, every day, from now on, almost forever. And data continues to grow. So 1PB today is likely to be 2PB of more in 12 months (it’s great to be in the storage business). 

So those are the challenges. How can it be done, effectively, day in and day out, enough so that IT can depend on their data being backed up.

Igneous to the rescue…

First, Igneous came out of stealth a while back (listen to our podcast) with a couple of unique capabilities needed for massive data repository discovery and analysis. That is they built a unique engine to scan and index PB scale data repositories. This was so they couldd provide administrators better visibility into their PB scale data repositories. But this isn’t about that product, it’s about backup. 

But some of the capabilities they needed to support that product helped them perform backups as well. For instance, their scan needed to handle PBs of data. They came up with AdaptiveSCAN, which didn’t use standard NFS and SMB data transfer protocols to gain access to file metadata. To open a file on NFS or SMB takes quite a lot of NFS or SMB transactions. But to access metadata only, one doesn’t have to use all those NFS and SMB capabilities, it can be done with much less overhead even when using NFS or SMB.

Of course having a way to scan Billions of files was a major accomplishment, but then where do you put all that metadata. And how can you access it effectively to support backup up a PB data repository. So they needed some serious data indexing capabilities and so came up with InfiniteINDEX

Now a trillion item index, seems a bit much, even for PB scale repositories. But my guess is they have eyes on taking their PB scale backups and going after even bigger fish,. That is offering backups for EB scale data repository. And that might just take a trillion item index

Next, there’s moving PB or even TB of data quickly is no small trick. As the development team at Igneous mostly came from unstructured data providers, they also understood and have access to APIs for most storage vendors (NetApp, Dell-EMC Isilon, Pure FlashBlade, Qumulo, etc.). As such, where available, they utilized those native vendor storage API calls to help them move data rather than having to Open an NFS or SMB file and Read it. 

Of course, even doing all that, moving 100TBs of data around or scanning PB sized data repositories is going to take a lot of processing and IO bandwidth to do in a reasonable period of time. 

So another capability they developed is massive parallelism. That is being able to distribute scan, indexing or data movement work, out to multiple systems. In that fashion it can be accomplished in significantly less wall clock time. 

Well with all that, they pretty much had the guts of a backup application system for PB data repositories but they still didn’t have the glue to put it all together. But recently they announced just that a Igneous’s DataProtect, a full scale backup application for PB of data. 

I suppose I haven’t done justice to all of what they have developed or talked about at their session, so I would suggest viewing their talk at CFD8 and listening to our GBoS podcast to learn more. They did demo their product at CFD8 but I believe it was a canned demo.

I didn’t think I’d see the day when some vendor would offer backup services for PBs of data let alone be shooting for more, but there you have it. Igneous means to take your PB scale data repositories and make them as easy to operate as TB scale data repositories. They call that democratizing data.


See these other CFD8 bloggers write ups on Igneous.

CFD8  – Igneous Follow Up  by Nate Avery (@Nathaniel_Avery)

Picture credit(s): All from screen saves during Igneous’s session at CFD8