BioIT World 2016 - HPC Trends from the Trenches

1
Trends from The Trenches
2016 BioIT World Conference & Expo
Photo credit: Aaron Gardner

2
I’m Chris.
I’m an infrastructure geek
I work for the BioTeam.
Photo credit: Cindy Jessel

3
www.BioTeam.net
Independent Consulting Shop
Run by scientists forced to learn IT to “get science done”
Virtual company with nationwide staﬀ
15+ years “bridging the gap” between hardcore science, HPC & IT
Honest. Objective. Vendor & Technology Agnostic.

4
BioTeam in ‘16 
missing: Cindy & Mark

6
Been doing this stuﬀ since 2002
Midlife crisis time …

7
Worried that this talk is becoming
tired or too repetitive. Maybe we
need to change the format and/or
speaker.
6-question online survey is here:  
http://biote.am/survey16
This scannable QR code  
links to the survey

8
People doing IT for life science must
embrace the existential fear, dread
and uncertainty.
{ this is a constant; not a trend }

9
Bottom Line: 
Science evolving faster than IT can refresh
infrastructure, design patterns & operational
practices

10
We have to build stuﬀ that runs for years
… to support researchers who can’t
articulate their needs beyond a few months

11
Trends: DevOps, Automation & Career Stuﬀ

12
This section 80% recycled from ’15 talk
… thought about skipping the old sides 
but the topic is too important

13
SysAdmins: If you can’t script your
upward career mobility is done. Period.

14
It’s not “just” cloud & virtualization …
Everything will have an API soon

15
Orchestration, conﬁguration
management & automation stacks are
already in your enterprise

16
You will need to learn them.
Bonus: You will be 10x more productive
Bonus: Your LinkedIn proﬁle views will go up 300%

17
Organizations: “We are not on the
cloud yet” is no longer a viable excuse

18
DevOps methods & infrastructure
automation have been transforming
on-premise IT for years at this point

19
These methods are transformational
“force-multipliers” for overworked &
understaﬀed IT teams

20
Not using these methods today implies
a certain ‘legacy’ attitude or method of
operation

21
.. that your competition may not have

22
Chef, Puppet, Ansible, SaltStack …
Pick what works for you (and that all can agree on)
… and commit to learning, using and evangelizing it
… ideally across the enterprise (not just Research)

23
Chef, Puppet, Ansible, SaltStack …
Pick what works for you (and that all can agree on)
… and commit to learning, using and evangelizing it
… ideally across the enterprise (not just Research)
… and if you stop by our booth ask to see the awesome way we are
doing automated and simple appliance deployments and updates via
USB wireless dongle, iphone/tablet & automation magic …
Happy to chat/demo this stuﬀ

24
Hey Network Engineers …
Same API-driven automation trends are steamrolling your way
You can’t coast on a Cisco certiﬁcation for much longer

25
New for 2016: Bench Scientists …
‘big data’ is becoming mainstream. Spreadsheets and Excel
macros are not going to be viable for some of you very quickly.
May have to learn new data analytic techniques to get the best
info out of your most interesting data

26
New for 2016: Data Scientist Roles
Sexy new occupation. High demand/salaries across many industries 
Warning: We may be seeing signs of rot in the talent pool similar to
‘bioinformatician’ craze in the ‘90s …
Blunt truth: Analytic mastery requires BOTH skill AND domain
expertise. Watch out for certiﬁcate/diploma mills. You don’t need
the mythical “unicorn” but you need people who understand the
domain and not just the toolset.

27
New for 2016: Data Engineer Role
[Emerging Trend ]  
Not so sexy and not near as much hype as ‘data scientist’
… but becoming increasingly essential in organizations with high-scale,
high-velocity data. Especially with diverse data types.
Blunt truth: Specialist skills are needed to manage data ingest,
provenance, transformation, movement and manipulation. It’s a waste of
resources to have a scientist do this and Enterprise IT does not have the
skillset.

28
Trends: Compute
Just the new/emerging stuﬀ …

29
Emerging compute pain / capability gap
Bench scientists losing out on workforce mobility initiatives
because they are‘tethered’ to labside analysis workstations

30
The issue:
‣ More and more lab side instruments require large
workstations to process results and look at data
• These users do not need HPC and may have had little
interaction with the ‘big compute’ people
• … but the hardware requirements for the vendor
software far exceed what their personal workstation is
capable of running
• Corporate VM / VDI teams have no idea

31
The result:
‣ Highly skilled, highly recruited and highly-paid PhD level
scientists are ‘tethered’ to wet lab workstations for many
hours a day
• … they CAN’T work from their oﬃce/desk
• … they CAN’T work remotely via VPN or VDI
‣ Productivity and morale issue today. Potential to become
an employee retention problem in the future

33
Compute Trend:
HPC Across the Enterprise + HPC as a Service

34
HPC for Enterprise; Not Just Research
‣ Recurring trend in recent HPC assessment projects
‣ Biotech, Pharma and Government
‣ HPC now rebuilt, rebranded, operated and advertised
as a service consumable by anyone in the organization
‣ Available via traditional methods
‣ … but also via exposed APIs that enable HPC to be
integrated in various upstream workﬂows
‣ Stretching our Infrastructure & Ops in interesting ways

35
HPC for Enterprise; Not Just Research
‣ Requires significant investment in user outreach,
onboarding, training, knowledge transfer & mentoring
‣ Increased focus on portals and common workflow
templates (cli via SSH is not for everyone …)
‣ For APIs we are seeing:
‣ Homegrown, Univa and http://agaveapi.co
‣ This also affects hardware infrastructure and
operations model a little bit …

36
‣ History Time:
‣ Discovery oriented research shops often intentionally
sacriﬁce high availability to save vast piles of money
‣ It may be annoying but we can tolerate downtime in
exchange for bigger/faster/better or more capable …
‣ This approach does not translate well into the
Enterprise world with SLAs and 24x7 uptime needs
‣ Lets look at one real world example …

37
See anything interesting here?

38
12x 40Gbps Uplinks (!)
48x 10Gbps ports ( for science! )
OMG! Full path diversity & redundancy via 2nd switchset

39
Common/Reconﬁgurable Hardware for HPC and Hadoop
Compute Trend:

40
[ Remember the ‘meta’ problem] — IT has to design infrastructures that
will live for years in face of unknowable scientific requirement 
“Big Data” may still be 70% vapid hype and vendor lies but the 30% that
is actually useful is showing excellent promise within the groups that
are deploying it.
HDFS fundamentally wants lots of local disk spindles as close as
possible to the CPU. This is starting to significantly affect the types of
servers that forward-looking BioIT people are procuring.
Compute 2: Common Kit for HPC/Hadoop

41
2U size with 24 mappable drives up front …
HPE Apollo

42
1
2
3
4
1
2
3
4
And 4x very dense servers packed into the rear
This is an interesting packaging mix, particularly for those seeking to support
‘traditional’ HPC and Hadoop/BigData on common hardware.

43
Compute Trend & Emerging Pain Point:
Data Lakes & Multi-tennant secure Hadoop

44
Compute 3: Hadoop 2.0 & Data Lakes
‣ Signiﬁcant Trend in 2016
‣ Feels like we’ve passed the 50% threshold of clients currently
running real workloads on “big data” platforms
‣ And just like HPC, these platforms are (often) being marketed
beyond just the research organization
‣ Interesting: Roughly 70:30 split between cloud and onsite stacks
‣ Signiﬁcant potential for wasted resources
‣ Hype may drive adoption in absence of realistic need
‣ Approach with caution (and with the right skills!)

45
Compute 3: Hadoop 2.0 & Data Lakes
‣ [Longstanding Issue] Just like “grid computing” in the 90s we are still in
the “empty hype”, “vapid press release” and “endless lies from marketing”
phase of Hadoop/BigData/DataLake
‣ [Blunt Advice] Despite vendor claims, this stuﬀ is currently neither easy,
simple or approachable by mere mortals.
‣ Not magic, you just have to work for it. Unicorns/ninjas not required.
‣ Getting useful, actionable results out of these systems REQUIRES a user
that is knowledgeable of the domain, data types, data structure,
common data transformations, SQL, statistics, visualization, etc. etc.
etc.
‣ Pre-sales people are often not truthful about this (!)

46
Apache Hadoop Ecosystem
‣ https://projects.apache.org/projects.html?category
‣ Fantastic open source building blocks with multiple
commercial & free options
‣ Two quick anecdotes / war stories
YARN, Spark, Hive, Tez, Pig, Zookeeper, HDFS, Ambari, etc. etc.

47
‣ Apache Hadoop Ecosystem consists of MANY independent open
source projects
‣ There are a fantastic number of moving parts that need consideration
when you are trying to build a functional and integrated stack
‣ There is a lot of coordination but each of these projects tends to have
varying levels of developer support, release cycles and maturity
‣ One example: Apache Zeppelin
‣ Web based notebook for interactive data analytics
‣ https://zeppelin.incubator.apache.org/
‣ Incredibly approachable very useful - end users LOVE it
Adventures with Zeppelin

49
‣ Almost every data analyst who sees Zeppelin loves it
‣ May become a standard for web-based ‘data science
notebooks’
‣ However … it’s still in “incubator” stage within
Apache Foundation
‣ And this presents some challenges with deployment
‣ User interest outpaces development roadmap

50
‣ When we 1st tried to deploy zeppelin for users …
‣ It had no integrated support for authentication or more than 1 user
‣ All notebooks, all data and all results were accessible and editable by
anyone with access to the (non secured, non-authenticated) web interface
‣ So we had to install multiple zeppelin servers
‣ Each user gets a private, customized install
‣ And it was our responsibility to create and enforce access controls
‣ Honestly not a hard issue to resolve
‣ But a good example of “interest outpacing development” in this space

51
‣ Deployment is smooth with no sensitive data or a single-user workflow
‣ Cloud environments are great as you can spin up bespoke clusters
‣ Bit harder when Enterprise SecOps compliance checkboxes are involved:
‣ Multi-user / Active Directory & Kerberos integration
‣ Encryption everywhere (at-rest, in-flight) for data and communication
‣ Role based access control for services and HDFS filesystem contents
‣ Audit trails & compliance/governance testing
‣ Authenticated API calls and secure API gateway
‣ Not all ecosystem components support hardened operation
‣ May be forced to remove elements (Hive, etc.) that can’t play in the hardened environment
Securing for Enterprise Use

52
Trends: Storage
Some new things and a few constant pain points

53
‣ Massive text files
‣ Massive binary files
‣ Flatfile ‘databases’
‣ Spreadsheets everywhere
‣ Directories w/ 6 million files
‣ Large files: 600GB+
‣ Small files: 30kb or smaller
This is what makes life science an interesting market
Still have ALL the file types
‣ This is
problematic
because many
storage products
are engineered to
to excel at only
1-2 of these
things …

54
‣ Organizations with 1PB+ or more data under
management that DO NOT have a full-time
employee managing/curating/monitoring things
‣ … are wasting more in storage CapEx than the fully
loaded cost of the human resource
Data Curation Still Important

55
‣ Highly skilled, highly compensated PhD’s still wasting
hours or days moving data around
‣ Easy to talk about; Hard to implement sensibly
‣ Operational timesink for human staff
‣ SOPs for manual ingest often non-existent
‣ Does IT even know when this occurs?
‣ Are scientists doing MD5 checksums each time a ﬁle moves
from location A to B?
‣ Network based data movement still safer/superior
Shipping disk drives fraught with risk
Data Ingest Still Risky & Hard

56
‣ Cloud storage cannot be ignored even by organizations
committed to a 100% on-premise footprint
‣ Even at institutions where space/power/cooling costs are
hidden or “free” to research users
‣ Amazon S3 and similar have become important:
‣ For data ingest and data exchange
‣ Long term data dissemination via “downloader pays” model
Cloud Storage Strategy Still Required

57
Storage: Few trends becoming more real in ‘16

58
‣ The friction required to ACQUIRE or GENERATE vast
amounts of scientific data is LOWER than it has ever
been
‣ Far easier to acquire data than to store/manage it
‣ Petabytes of open-access data available via Internet
and legit reasons why a researcher may want to bring
significant portions in-house for analysis
‣ We need governance in place now; Research users
must get used to providing biz/science justification for
storage consumption
Reality #1 - Ease of Generation/Access

59
‣ We have lost the battle to centralize large-scale data
providers and data consumers in managed locations
‣ Data-intensive instruments will continue to diﬀuse
throughout EVERY part of the organization
Reality #2 - Data Diﬀusion

60
‣ POCs are turning into procurement
‣ Single product tackles multiple pain points
1. Make cheap storage fast so I don’t have to buy the spendy stuff
2. Multi-vendor NAS behind a single consistent namespace
3. Multi-vendor NAS-to-NAS data migration
4. POSIX interface for local object storage
5. POSIX interface for multi-cloud object storage
6. Multi-cloud replication/sync for HPC “cloud bursting”
Reality #3 - Avere Systems

61
Storage: New/Interesting in 2016

62
Image source: https://csc.fi/image/blogs/entry?img_id=252729
Storage Related Current Event:
Great post-mortem csc.fi writeup on a
massive Lustre storage crash where 1.7
petabytes / 850 million files were lost
before being heroically (mostly) recovered
Short URL: http://biote.am/bq

63
‣ Huge install base of 3-5 year old Scale-out NAS
‣ … all approaching refresh/replace cycle time
‣ BioTeam alone has 4+ large-scale NAS re-assessment
projects going on right now - all options open
‣ Drivers coming from all sorts of directions
‣ Object, cloud, commodity SDS, disruptive kit, etc.
‣ 2016 could see a massive shift in the on-premise
storage landscape as organizations reevaluate storage
plaforms, plans and architectures
2016 is going to be a VERY INTERESTING year
New: Massive NAS Refresh/Replace

64
‣ Formally novel/disruptive stuﬀ is gaining acceptance
‣ New crop of scale-out NAS players (Qumulo, etc.)
‣ Software-based options (Scality, Rozo, etc.)
‣ Next-gen object store options (Igneous, etc.)
‣ Flash/SSD pricing becoming viable in our space
‣ And the DIY folks keep doing cool things
‣ Example: Pineda lab doing 2PB+ Lustre-on-ZFS green
storage at prices lower than the big cloud providers
‣ Link: http://biote.am/be
Very interesting implications for 2016+ storage refresh projects
New: Storage landscape getting interesting

65
‣STORAGE = CHEAP
‣METADATA = PRECIOUS
Critical topic when storage refresh projects being considered
New: “Data about our Data” is now essential

66
Quick Detour 
A few war stories from the “what the heck are
we keeping on these systems” battlefront …

About
67
‣ Core storage: Multi-PB EMC Isilon
‣ Method: Robinhood 3.0-alpha1 on a dedicated server
with MariaDB on XFS as the database. Full ﬁlesystem
walk via NFSv3 and auto mount
‣ Produced:
‣ Took 7.8 days to walk the ﬁlesystem and index 283 million
entries. Final database size was 175GB
‣ Re-indexing takes LONGER due to database interactions
Metrics #1: Robinhood w/ EMC Isilon

Basic findings
68
‣ 273 million files
‣ 1.29 PB total space consumed
‣ 5.07 MB average file size
‣ 3.64 million directories
‣ 321,000 symbolic links

Basic ﬁndings, continued
69
‣ Top User owned 327 TB ( ~28% of all data!)
‣ Top 10 Users responsible for ~93 % of all data
‣ Only about 30TB of 1.2PB is really active
‣ Majority of data written is NEVER read
‣ Raw instrument data
‣ Copies of public and partner data

Hard lessons learned
70
‣ Gathering this data was hard work but it may drive a
fundamental change in the enterprise scientiﬁc storage
platform
‣ Concrete evidence that scientists are using expensive
Tier-1 storage as a dumping ground for ﬁles that are
never looked at again
‣ A ~30TB active data size has interesting implications for
future architectures and potential use of SSD/Flash

An interesting noSQL approach to the same problem
71
‣ Can’t share results as this work is ongoing right now
‣ This is a group w/ ~5PB of Isilon and an Avere based
client access layer
‣ This group had previously tried ﬁlesystem scanning and
SQL data stores but hit the wall at 500 million records
‣ So they are doing something new …
Method #2: pwalk + redis vs Isilon/Avere

Basic Method
72
‣ Scan phase:
‣ Filesystem heiarchy broken up into walkable segments
‣ Individual pwalk tasks submitted as normal HPC jobs
‣ Manual tuning required as directories vary by size and depth
‣ Each pwalk job writes to a unique log ﬁle

Basic Method
73
‣ Load phase:
‣ pwalk jobs are read in parallel and sent to redis noSQL store
‣ Work is chunked to keep redis pegged across SMP cores
‣ Primary key = full path to ﬁlesystem object/directory/pointer
‣ Record = ﬁle stats + hash map of parent/child relationships

Basic Method
74
‣ Analysis phase:
‣ Massive secondary processing within redis to generate vast
new sets of keypair:record entries
‣ Effectively pre-computing the answer to all desired storage
metric questions

Basic Method
75
‣ Future Optimization:
‣ Avere exposes some pretty useful APIs
‣ Current idea is to use Avere API to report on which part of
the big filesystem is experiencing heavy activity
‣ … then spawn pwalk jobs that JUST traverse the recently
active files and folders
‣ … avoiding or reducing the need for full filesystem crawls

76
‣ Crawling ﬁlesystems is painful or not viable above a
certain scale or data churn level
‣ Expect to see more of:
‣ Storage platforms that offer native insight/analytics
‣ Storage platforms continuing to expose useful APIs
‣ Expect commercial oﬀerings as well
‣ BioTeam will be looking at ClarityNow! from
dataframeworks.com in 2016
Getting better “data about our data” - Future

77
I’m still saying this: 
 
Object storage is the future of
scientiﬁc data at rest.  
 
Period.

78
Storage: Object Is the Future
This what my metadata looks like on a POSIX ﬁlesystem:
Owner
Group membership
Read/write/execute permissions based on Owner or Group
File size
Creation/Modification/Last-access Timestamps

79
Storage: Object Is the Future
This is what I WOULD LIKE TO TRACK on a PER-FILE basis:
What instrument produced this data?  
What funding source paid to produce this data?  
What revision was the instrument/flowcell at?  
Who is the primary PI or owner of this data? Secondary? 
What protocol was used to prepare the sample?  
Where did the sample come from?  
Where is the consent information? 
Can this data be used to identify an individual?  
What is the data retention classification for this file?  
What is the security classification for this file?  
Can this file be moved offsite?
etc. etc. etc.
…

80
Storage: Object Predictions
‣ It will still be many years (and a long transition) before we are
native object for scientiﬁc data at rest (you’ve got time!)
‣ Storage refresh projects will drive testing and POCs in 2016+
‣ Metadata, cost and erasure coding will be the primary drivers
‣ Expect to see
‣ A lot of POSIX/Object gateways and transition products
‣ Data or image-heavy commercial software platforms will
begin to support native object stores

81
Trends: Cloud
A few new things …

82
Cloud: Recap
‣ Core drivers behind cloud adoption remain unchanged
‣ Remember, it’s about CAPABILITY not COST for many
‣ Cloud footprints being designed for security, data exchange
and multi-party collaboration
‣ In 2015-16 BioTeam has seen a large increase in
‣ Hybrid cloud architectures
‣ Multi-cloud design patterns and architectures
‣ Clients procuring direct cloud connectivity (1Gbps, 10Gbps)

83
Cloud: New for 2016
‣ BioTeam is predicting a new era of CLOUD SOBRIETY
‣ We expect to see a few major cloud pullbacks in 2016-2017
‣ Root cause:
‣ C-level executives ordering “cloud ﬁrst” without doing the math
‣ Long-time cloud users now have actionable cost & spend data
‣ Commodity price trends for on-premise kit looking more
interesting when cloud storage bill exceeds $100,000/month

84
Cloud: New for 2016
‣ The cloud pullback/clawback projects will be interesting
‣ Some of them may go elsewhere rather than back to on-premise
‣ Interesting movements within the traditional large supercomputing
centers. Some oﬀering services to Industry in order to obtain new
revenue/funding/cost-recovery streams
‣ 2016-2017 may see workloads pulled from IaaS cloud platforms
and redeployed at places like NCSA iForge rather than on-premise
or in a standard colo suite

85
Cloud: New for 2016
‣ Starting to see interesting new criteria for Cloud vendor selection
‣ Now you have to ask these questions:
‣ “What are the chances that my cloud provider is going to start a
pharmaceutical or healthcare company that will directly compete with
me?”
‣ “Are product teams monitoring my API usage? Who are they
communicating with? Do they use this information to make my life
easier or are they collecting competitive intelligence that may be used
to compete against me in the future?”
‣ My cynical (personal) opinion: Cloud providers must be screened for
competitive risk in 2016 and beyond. I’m getting concerned & suspicious.

87
Lots of recycled slides from last year
Keeping them because it’s the most
important section of the talk.
This really is going to be our hardest
challenge over the next few years

88
A conﬂuence of unfortunate events …
• Terabyte-scale wet lab instruments are everywhere
• Centralizing high-data-rate instruments = umpossible :)
• Petabytes of interesting open-access data on Internet
• Petabytes of interesting data w/ collaborators
• Tera|Petabyte scale “data spread” seems unavoidable
• Known: data will span sites, buildings and clouds

89
Data Intensive Science Will Require
• Ludicrous bandwidth @ the network core
• Switches & routers that don’t drop packets
• Massive bandwidth delivered to Top of Rack
• Very fast (40g or 100g) to campus buildings
• Very fast (10g or 40g) to lab spaces
• 1Gbps or 10Gbps Internet/Internet2 access
• Security controls that can handle single-stream 10Gbps
+ data ﬂows without melting down

91
Terabyte-scale data movement is
going to be an informatics “grand
challenge” for the next 2-3+ years
And far harder/scarier than previous compute & storage challenges

92
Issue #1
Current LAN/WAN stacks bad for emerging use case
Existing technology we’ve used for decades has been architected to
support many small network ﬂows; not a single ‘elephant’ data ﬂow

93
Issue #2
Ratio of LAN:WAN bandwidth is out of whack
We will need faster links to “outside” than most organizations have
anticipated or accounted for in long-term technology planning

94
Issue #3
Core, Campus, Edge and “Top of Rack” bandwidth
We need bandwidth everywhere; Enterprise networking caught by
surprise. None of this stuﬀ (100 Gig, 40 Gig outside of the datacenter,
etc.) is on their long-term roadmap (or budgeted)

95
Issue #4
Bigger blast radius when stuﬀ goes wrong
Compute & storage can be logically or physically contained to minimize
disruption/risk when Research does stupid things.
 
Networks, however, touch EVERYTHING EVERYWHERE. Major risk.

96
Why this will be diﬃcult to achieve

97
Issue #5
Social, trust & cultural issues
We lack the multi-year relationship and track record we’ve built with
facility, compute & storage teams. We are “strangers” to many WAN and
SecurityOps types

98
Issue #6
Our “deep bench” of internal expertise is lacking
Research IT usually has very good “shadow IT” skills but we don’t have
homegrown experts in BGP, Firewalls, Dark Fiber, Routing etc.

99
Issue #7
Cost. Cost. Cost.
Have you seen what Cisco charges for a 100Gbps line card?

100
Issue #8
Firewalls, SecOps & Incumbent Vendors
Legacy security products supporting 10Gbps can cost $150,000+ and
still utterly fail to perform without heroic tuning & deep conﬁg magic.
Alternatives exist but massive institutional inertia to overcome.  
 
Deeply Challenging Issue.

Gulp. 
So what do we do?
101

102
‣ Acknowledge that we are entering a new
chapter in an era of “data intensive science”
‣ Understand that we now have high scale data
producers and consumers diﬀusing everywhere
‣ … Including clouds, partners and collaborators
‣ Recognize that new methods are needed and
that this may include new vendors/platforms

103
‣ Engage ASAP with enterprise networking, WAN and
security teams - they now need to be part of the scientiﬁc
computing family
‣ Survey networking roadmap, budget and capabilities
‣ Historically network engineering groups are under-
resourced relative to Research
‣ [Research Leadership] If you want to make new friends in
networking, become the outside champion arguing that
they need more resources
Advice #1

104
‣ Survey your own labs, ﬂoors and buildings to document the
large-scale data producers and consumers
‣ Talk to PIs about their instrument procurement plans
‣ Talk to business leadership to understand collaboration
eﬀorts and use of external partners or data providers
‣ Actively seek out the folks moving data on physical disks
Advice #2

106
Internet2 Is Open for Business
10Gig Internet2
link @ pharma client

107
‣ Did you see Dan’s talk yesterday?
‣ Since 2015 we have seen signiﬁcant increase in # of BioTeam
clients choosing to make use of Internet2 for high speed access
to collaborators and clouds
‣ At least ~6 now spanning commercial and US.gov
‣ Pro: 100 Gig network purpose built for data intensive science
‣ Con: I2 is not an ISP so connecting is non-trivial, especially if you
want Layer-3 access and don't have a /24 public IPv4 block to
spare
Internet2 and High Speed R&E Networks

108
And this brings us to …
ScienceDMZ

109
‣ The “data intensive science” era may ﬁnally
trigger a process by which organizations
logically or physically separate the “business
network” from the “science/data” network
‣ Huge deal. This is a fundamental shift in
practices that have largely been static for
decades
ScienceDMZ #1

110
‣ Other ﬁelds (High Energy Physics, etc.) have faced
this problem before and solved it
‣ The philosophy, concepts and methods they use have
been broadly grouped under the descriptive term
“ScienceDMZ”
‣ Evangelized by nice folks at  
http://fasterdata.es.net
‣ In use today in large scale production environments.
No BS. It works.
ScienceDMZ #2

111
How to sell ScienceDMZ in 1 image

112
‣ Keep large ﬂows oﬀ of business networks
‣ Build TCP networks that do not drop packets
‣ Instrument at multiple locations via perfSONAR
‣ Use grown-up xfer tools, not FTP, RSYNC or SCP
‣ Completely rethink how network security and
intrusion detection will be performed
Practical ScienceDMZ Elements

113
‣ Our second dedicated Converged IT summit will be
in San Diego this year - October 24-26, 2016
‣ We have a 100 Gig test lab hosted by TACC in
Austin, TX and supported by many of our vendor
friends - including a 100 gig Internet2 link
‣ We intend to build, test, benchmark and talk about
various types of ScienceDMZ network designs
‣ We intend to share this info freely and (ideally) open
our lab up for others to use/test/research
What BioTeam is doing

114
Happening right this instant …
BioTeam + Intel + Aspera + Internet2 100 Gbps data transfer trials

115
‣ BioTeam 100 Gig testbed @ TACC in Austin, TX
‣ 2x Intel servers at each end of the Internet2
VLAN, each with 3x 40GB NICs
‣ 100 Gig circuit from TACC to NSCA via Internet2
‣ Aspera data-xfer software + Intel DPDK tech
‣ Goal: See how fast we can move real data
server-to-server across a 100 Gig link

119
‣ Current status
‣ Still sorting out hardware issues and tuning the OS,
NIC and driver stack for best possible throughput
‣ Currently seeing 73Gbit/sec Memory-To-Memory
‣ Will do disk-to-disk at BioIT’16 conference this week
‣ Very interesting process and some unexpected results
‣ Thermal management is apparently a big deal
‣ Testing teased out non-trivial hardware issues: disk,
PCI risers, NICs and multiple motherboards

120
end; Thanks!
slideshare.net/chrisdag/ chris@bioteam.net @chris_dag
This scannable QR code  
links to the survey
6-question online survey is here:  
http://biote.am/survey16

BioIT World 2016 - HPC Trends from the Trenches

More Related Content

What's hot (20)

Viewers also liked (13)

Similar to BioIT World 2016 - HPC Trends from the Trenches (20)

More from Chris Dagdigian (8)

Recently uploaded (20)

BioIT World 2016 - HPC Trends from the Trenches