Repo is here: https://github.com/joelandman/rzf. There’s a lightning talk to go along with it, and I’ll make sure I can get it together for this as well.

I’ve not heard of aria2c before today. Sort of a super wget as far as I could tell. Does parallel transfers to reduce data motion time, if possible. So I pulled it down, built it. I have some large data sets to move. And a nice storage area for them. Ok. Fire it up to pull down a 2GB file. Much faster than wget on the same system over the same network.

Nice machine we have here … root@hermes:/data/tests# lspci | egrep -i '(AMD|NVidia)' | grep VGA 3b:00.0 VGA compatible controller: <a href="http://www.pny.com/nvidia-quadro-gp100">NVIDIA Corporation GP100GL</a> (rev a1) 88:00.0 VGA compatible controller: Advanced Micro Devices, Inc. [AMD/ATI] <a href="http://www.tomshardware.com/reviews/amd-radeon-vega-frontier-edition-16gb,5128.html">Vega 10 XTX</a> [Radeon Vega Frontier Edition] I want to see how tensorflow and many others run on each of the cards. The processor is no slouch either: root@hermes:/data/tests# lscpu | grep "Model name" Model name: Intel(R) Xeon(R) Gold 6134 CPU @ 3.

There are a set of modules in the Perl universe that I’ve been looking for an excuse to use for a while. They are the MCE set of modules, which purportedly enable easy concurrency and parallelism, exploiting many core CPUs, and a number of techniques. Sure enough, I had a task to handle recently that required this. I looked at many alternatives, and played with a few, including Parallel::Queue. I thought of writing my own with IPC::Run as I was already using it in the project, but I didn’t want to lose focus on the mission, and re-invent a wheel that already existed elsewhere.

Though, when I look at the “great speed”, it is often on par with or less than Scalable Informatics sustained years before. From 2013 SC13 show, on the show floor, after blasting through a POC at unheard of speed, and setting long standing records in the STAC-M3 benchmarks … Article in question is in the Register. Some of the speeds and feeds: * 200 microsecs latency * 45GBps read bandwidth * 15GBps write bandwidth * 7 million IOPS But then … a fibre connection.

Way way back in the early aughts (2000’s), we had built a set of designs for an accelerator system to speed up things like BLAST, HMMer, and other codes. We were told that no one would buy such things, as the software layer was good enough and people didn’t want black boxes. This was part of an overall accelerator strategy that we had put together at the time, and were seeking to raise capital to build.

For a partner. They made a request for something we’ve not built in a while … it had been end of lifed. One of our old Pegasus units. A portable deskside supercomputer. In this case, a deskside franken-computer … built out of the spare parts from other units in our lab. It started out as a 24 core monster, but we had a power supply burn out, and take the motherboard with it.

While the day job builds (hyperconverged) appliances for big data analytics and storage, our partners build the tools that enable users to work easily with astounding quantities of data, and do so very rapidly, and without a great deal of code. I’ve always been amazed at the raw power in this tool. Think of a concise functional/vector language, coupled tightly to a SQL database. Its not quite an exact description, have a look at Kx’s website for a more accurate one.

Most people use squashfs which creates a read-only (immutable) boot environment. Nothing wrong with this, but this forces you to have an overlay file system if you want to write. Which complicates things … not to mention when you overwrite too much, and run out of available inodes on the overlayfs. Then your file system becomes “invalid” and Bad-Things-Happen(™). At the day job, we try to run as many of our systems out of ram disks as we can.

TL;DR version: 10GB/s write, 10GB/s read in a single 2U unit over 100Gb network to a backing file system. This is tremendous. The system and clients are using our default tuning/config. Real hyperconvergence requires hardware that can move bits to/from storage/networking very quickly. This is that. These units are available. Now. In volume. And are very reasonably priced (starting at $1USD/GB). Contact us for more details. This is with a file system …

I am trying to help people conceptualize the experience. Here is a video depicting very fast, very powerful cars and their sound signatures. This is a good start. Take one of those awesome machines, and turn off half the engine. So it is literally running with 1/2 of its power turned off. Remember this. There will be a quiz. As we flippantly noted in the video, this is face-melting performance. Had I any hair left, it would have been blown way back.

We’ve got customers with very large resource machines. And software that sees all those resources and goes “gimme!!!!”. So people run. And then more people use it. And more runs. Until the resources are exhausted. And hilarity (of the bad kind) ensues. These are firedrills. I get an open ticket that “there must be something wrong with the hardware”, when I see all the messages in console logs being pulled in from ICL saying “zOMG I am out of ram ….

Simple test case, no file system … using raw devices, what can I push out to all 60 drives in 128k chunks. Actually this is part of our burn-in test series, I am looking for failures/performance anomalies. ----total-cpu-usage---- -dsk/total- -net/total- ---paging-- ---system-- usr sys idl wai hiq siq| read writ| recv send| in out | int csw 0 1 95 5 0 0| 513M 0 | 480B 0 | 0 0 | 10k 20k 4 2 94 0 0 0| 0 0 | 480B 0 | 0 0 |5238 721 0 2 98 0 0 0| 0 0 | 480B 0 | 0 0 |4913 352 0 2 98 0 0 0| 0 0 | 570B 90B| 0 0 |4966 613 0 2 98 0 0 0| 0 0 | 480B 0 | 0 0 |4912 413 0 2 98 0 0 0| 0 0 | 584B 92B| 0 0 |4965 334 0 2 98 0 0 0| 0 0 | 480B 0 | 0 0 |4914 306 0 2 98 0 0 0| 0 0 | 636B 147B| 0 0 |4969 483 0 2 98 0 0 0| 0 0 | 570B 0 | 0 0 |4915 377 8 8 50 32 0 2|7520k 8382M| 578B 0 | 0 0 | 76k 215k 9 7 30 52 0 3|8332k 12G| 960B 132B| 0 0 | 109k 279k 10 5 29 53 0 2|4136k 12G| 240B 0 | 0 0 | 109k 277k 12 6 29 51 0 2|4208k 12G| 240B 0 | 0 0 | 108k 280k 11 6 31 50 0 2|2244k 12G| 330B 90B| 0 0 | 109k 281k 11 6 30 50 0 3|2272k 13G| 240B 0 | 0 0 | 110k 281k Writes around 12.

Oh no … we didn’t say “face melting” … did we? Oh. Yes. We. Did. The interview is here at the always wonderful InsideHPC.com You can see the video itself here on YouTube, but read Rich’s transcript. I was losing my voice, and he captured all of the interview in text. Take home messages: Insane IO/Networking/processing performance, small footprint, tiny price, available for orders now.

Brendan Gregg’s writings are always interesting and informative. I just saw a link on hacker news to a presentation he gave on “Broken Performance Tools”. It is wonderful, and succinctly explains many thing I’ve talked about here and elsewhere, but it goes far beyond what I’ve grumbled over. One of my favorite points in there is slide 83. “Most popular benchmarks are flawed” and a pointer to a paper (easy to google for).

I mentioned previously that we have some new (shiny) things … and it looks like you’ll be able to hear about them at my talk. See the schedule for timing information. This said, please note that we have a terrific line up of people giving talks: Fintan Quill from Kx on kdb+ … which is an awesome market leading Big Data Time Series analytics and database tool that runs absolutely balls-out insanely fast on our architecture Christian Mohrbacher from Thinkparq on BeeGFS … the primary parallel file system we are leveraging for Unison parallel file system appliances * Mark Nelson from Inktank/Red Hat on Ceph … the reliable block and object storage system that we’ve built into our Unison Object/Block Storage appliance * Doug Eadline from Basement Supercomputing on Hadoop, and likely showing a Limulus deskside Hadoop appliance * Phil Mucci from Minimal Metrics on optimization problems for systems and code.

This was a long time coming. Something I needed to do, in order to build a far better code capable of using less network, less CPU power, and providing a better overall system. In short, I ripped out the graphite bits and wrote a native interface to InfluxDB. This interface will also be adapted to kdb+ (32 bit edition), and graphite as time allows. In the process, I cleaned up a tremendous amount of code.

We are involved in a fairly large benchmark for a potential customer. I won’t go into many specifics, though I should note that lots of our Unison units are involved. Current architecture has 5 storage nodes (6th was temporarily removed to handle a customer issue). Each Unison node has a pair of 56GbE NICs, as well as our appliance OS, and bunches of other goodness (quite a bit of flash). Total capacity for test is of order 200TB of flash.

No, really mostly to clean up an accumulation of technical debt that was really bugging the heck out of me. I like Makefiles and I cannot lie. So I like encoding lots of things in them. But it wound up hardwiring a number of things that shouldn’t have been hardwired. And made the builds brittle. When you have 2 released/supported kernels, and a handful of experimental kernels, it gets hard making changes that will be properly reflected across the set.

I just saw this about doing a divide and conquer approach to massive scale genomics calculation. While not specific to the code in question, it looked familiar. Yeah, I think I’ve seen something like this before … and wrote the code to do it. It was called SGI GenomeCluster. It was original and innovative at the time, hiding the massively parallel nature of the computation behind a comfortable interface that end users already knew.

The Register has a great article on storage unicorns. Unicorns are not necessarily mythical creatures in this context, but very high valuation companies that appear to defy “standard” valuation norms, and hold onto their private status longer than those in the past. That is, they aren’t in a rush to IPO or get acquired. The article goes on to analyze the “storage” unicorns, those in the “storage” field. They admix storage, nosql, hyperconverged, and storage as a service.

A few years ago, we demonstrated some truly awesome capability in single racks and on single machines. We had one of our units (now at a customer site), specifically the unit that set all those STAC M3 records, showing this: and a rack of our units (now providing high performance cloud service at a customer site) for 8k random reads across 0.25 PB of storage on a very fast 40GbE backbone.

Should be fun, we will have booth (#461) on the side near the thoroughfare for the talks. Our HPC on Wall Street booth looked like this: [ ](/images/HPConWS-booth-spring2015.jpg) The display on the monitor is from our FastPath Cadence machine, and is part of the performance dashboard, built upon InfluxDB, Grafana, sios-metrics, and influxdbcli. Here is a blown up view, note the vertical axes for BW (GB/s) and IOPs. [ ](/images/cadence-dash-spring2015.jpg)

This is a very exciting hyper-converged system, representing our next generation of time series, and big data analytical systems. Tremendous internal bandwidths coupled with massive internal parallelism, and minimal latency design on networks. This unit has been designed to focus upon delivering the maximal performance possible in an as minimal footprint … both rack based and cost wise … as possible. You can use these as independent stand alone units, integrate them into a larger FastPath Unison system We have our software stack (SIOS) integrated onto each unit, and include our builds of Python + Pandas/SciPy/NumPy, R, and Perl.

Our Q1 is usually quiet, fairly low key. Not this one. Looks like lots of pent up demand. We are deep into record territory, running 200+% of normal, with possibility of more. Another new wrinkle is that our small investment round is mostly complete. This is new territory for us, and you may have noticed I’d backed off posting intensity over the last half year or so while this was going on.

I had hinted at this last week, so I figure I better finish working on this and get it posted already. The previous bit with language choice wakeup was about the cost of Foreign Function Interfaces, and how well they were implemented. For many years I had honestly not looked as closely at Python as I should have. I’ve done some work in it, but Perl has been my go-to language.

Our machines will be there, helping power the genomics pipelines to tremendous performance. Performance is an enabling feature. Without it you cannot even begin to hope to perform massive scale analytics. With it, you can dream impossible dreams. This article came out talking about a massive performance analytics pipeline at Nationwide Children’s Hospital in Ohio. This pipeline runs on a cluster attached to Scalable Informatics FastPath Unison storage. This is a very dense, very fast system, interconnected with Mellanox FDR Infiniband, Chelsio 40GbE, and leveraging BeeGFS from thinkparq.

So we build our own kernels. No great surprise, as we put our own patches in, our own drivers, etc. We have a nice build environment for RPMs and .debs. It works, quite well. Same source, same patches, same make file driving everything. We get shiny new and happy kernels out the back end, ready for regression/performance/stability testing. Works really well. But … but … parallel builds (e.g. leveraging more than 1 CPU) work only for the RPM builds.

Its that time of year, when the inventory fairies come out and begin their counting. Math isn’t hard, but the day job would like a faster and easier count this year. So, the day job is working on selling off existing inventory. We have 4 units ready to go out the door to anyone in need of 70-144TB usable storage at 5-6 GB/s per unit. Specs are as follows: 16-24 processor cores 128 GB RAM 48x {2,3,4} TB top mount drives 4x rear mount SSDs (OS/metadata cache) Scalable OS (Debian Wheezy based Linux OS) 3 year warranty As this is inventory reduction, the more inventory you take, the happier we are (and the less work that the inventory fairies have to do).

An interesting thing happened over the last few months and years. Systemd, a replacement init process for Linux, gained more adherents, and supplanted the older style init.d/rc scripting in use by many distributions. Ubuntu famously abandoned init.d style processing in favor of upstart and others in the past, and has been rolling over to systemd. Red Hat rolled over to Systemd. As have a number of others. Including, surprisingly, Debian. For those whom don’t know what this is, think of it this way.

Report XTR141111 was just released by STAC Research for the M3 benchmarks. We are absolutely thrilled, as some of our records were bested by newer versions of our hardware with newer software stack. Congratulations to Lucera, STAC Research for getting the results out, and the good folks at McObject for building the underlying database technology. This result continues and extends Scalable Informatics domination of the STAC M3 results. I’ll check to be sure, but I believe we are now the hardware side of most of the published records.

Here’s the basic theory behind swap space. Memory is expensive, disk is cheap. Only use the faster memory for active things, and aggressively swap out the less used things. This provides a virtual address space larger than physical/logical memory. Great, right? No. Heres why. swap makes the assumption that you can always write/read to persistent memory (disk/swap). It never assumes persistent memory could have a failure. Hence, if some amount of paged data on disk suddenly disappeared, well … Put another way, it increases your failure likelihood, by involving components with higher probability of failure into a pathway which assumes no failure.

This will be fun to watch run … Scalable Informatics FastPath Unison for the win!

Scalable’s booth is #3053. We’ll have some good stuff, demos, talks, and people there. And coffee. Gotta have the coffee. More soon, come by and visit us!

For Scalable Informatics Unison product. The whole system: [ ](/images/dash-2.png) Watching writes go by: [ ](/images/dash-3.png) Note the sustained 40+ GB/s. This is a single rack sinking this data, and no SSDs in the bulk data storage path. This dashboard is part of the day job’s FastPath product.

I’d like to offer an excuse. But I can’t. It was one single missing newline. Just one. Missing. Newline. I changed my config file to use port 10000. I set up an nc listener on the remote host. nc -k -l a.b.c.d 10000 Then I invoked the code. And the data showed up. Without a ()&(&%&$%*&(^ newline. That couldn’t possibly be it. Could it? No. Its way to freaking simple.

I tried the nodejs version and … well … it was horrible. Basic things didn’t work. Made life very annoying. So, being a good engineering type, I wrote my own. It will be up on our site soon. Here’s an example ./influxdb-cli.pl --host 192.168.5.117 --user test --pass test --db metrics metrics> \list series .----------------------------------. | series name | +----------------------------------+ | lightning.cpuload.avg1 | | lightning.cputotals.idle | | lightning.cputotals.irq | | lightning.

This has been a point of contention for us for years. Our office has multiple network attachments, using Comcast is part of it. This is the main office, not the home office. Latency on the link, as measured by DNS pings, have always been fairly high, in the multiple 2-3ms region, as compared to our other connection (using a different provider and a different technology) which has been consistently, 0.5ms for the last 2 years.

This is one of our engineering prototypes that we had to clear space for. A couple of new features I’ll talk about soon, but you should know that these are 12g SAS machines (will do 6g SATA of course as well). Front of unit: [ ](/images/IMG_2330.JPG) Note the new logo/hand bar. The rails are also brand new, and are set to enable easy slide in/out even with 100+ lbs of disk in them.

A 1.2PB single mount point Scalable Informatics Unison system, running an MPI job (io-bm) that just dumps data as fast as the little Infiniband FDR network will allow. Our test case. Write 2TB (2x overall system memory) to disk, across 48 procs. No SSDs in the primary storage. This is just spinning rust, in a single rack. This is performance pr0n, though safe for work. usn-01:/mnt/fhgfs/test # df -H /mnt/fhgfs/ Filesystem Size Used Avail Use% Mounted on fhgfs_nodev 1.

I am lucky. I fully admit this. There are people out there whom will tell you that its pure skill that they have been in business and been successful for a long time. Others will admit luck is part of it, but will again, pat themselves on the back for their intestinal fortitude. Few will say “I am lucky”. Which is a shame, as luck, timing (which you can never really, truly, control), and any number of other factors really are critical to one being able to have the luxury of doing what we are doing.

… and believe it or not, I forgot to mention it. This is the first time in company history that we had a backlog going into Q1. Orders being built and tested on the last work day of the year. We grew, not the amount we had originally forecast, but we understand why (and sadly have little control over that aspect). We are working very hard on our appliances … I am blown away as to how perfect a fit they are for folks.

We’ll be showing off 2 appliances, with a change of what we are showing/announcing on one due to something not being ready on the business side. The first one is our little 108 port siRouter box. Think ‘bloody fast NAT’ and SDN in general, you can run other virtual/bare metal apps atop it. The second will be a massive scale parallel SQL DB appliance. Usable for big data, hadoop like workloads, and other similar workloads more commonly used on other well known platforms.

See the Systems Engineering position here, and the System Build Technician position here. I’ll get these up on the InsideHPC.com site and a few others soon (tomorrow). But they are open now. For the Systems Engineering position, we really need someone in NYC area with a strong financial services background … Doug made me take out the “able to leap tall buildings in a single bound” line, as well as the “must be able to talk customers through complex vi sessions on system configuration files while driving 70 mph on a highway.

Oh yes we did. Oh yes. We did. This is the fastest storage box we are aware of, in market. This is so far outside of ram, and outside of OS and RAID level cache … [root@siFlash ~]# fio srt.fio ... Run status group 0 (all jobs): READ: io=786432MB, aggrb=23971MB/s, minb=23971MB/s, maxb=23971MB/s, mint=32808msec, maxt=32808msec This is 1TB read in 40 seconds or so. 1PB read in 40k seconds (1/2 a day).

I finally took some time from a busy schedule to play with AVX. I took my trusty old rzf code (Riemann Zeta function) and rewrote the time expensive inner loop in AVX primatives hooked to my C code. As a reminder, this code is a very simple sum reduction, and can be trivially parallelized (sum reduction). Vectorization isn’t as straightforward, and I found that compiler auto-vectorization doesn’t work well for it.

The IBM folks have turned the Blue Gene into what they claim is the worlds fastest blast engine. Interesting read. They use our A. thaliana data in the Bioinformatics Benchmark System v3 (BBS) to perform their measurement, as well as data from Aaron Darling for mpiBLAST. Our data had been in a mislabeled file for years, and I never took the time to rename the S. lycopersicum for the original Arabidopsis.