Common Design Elements for Data Movement Eli Dart
1 like554 views
The document discusses the challenges and strategies for high-performance data movement in data-intensive science, emphasizing the need for reliable data transfer between remote locations. It introduces the 'science DMZ' model, which incorporates dedicated resources and design patterns to optimize network performance and reliability for data transfers. The importance of standardization among data transfer tools and the role of reliable networks in enhancing data movement capabilities are also highlighted.
Common Design Elements for Data Movement Eli Dart
- 1. Common Design Elements for Data Movement Eli Dart, Network Engineer ESnet Science Engagement Lawrence Berkeley Na<onal Laboratory Cosmology CrossConnects Workshop Berkeley, CA February 11, 2015
- 2. Overview 2/10/15 2 • Context • Design paJerns • What do we need to do?
- 3. Context • Data-‐intensive science con<nues to need high-‐performance data movement between geographically distant loca<ons – Observa<on (or instrument) to analysis – Distribu<on of data products to users – Aggrega<on of data sets for analysis – Replica<on to archival storage • Move computa<on to data? Of course! Except when you can’t… – A liquid market in fungible compu<ng alloca<ons does not exist – Users get an alloca<on of <me on a specific compute resource – if the data isn’t there already, it needs to be put there – If data can’t be stored long-‐term where it’s generated, it must be moved – Other reasons too – the point is we have to be able to move Big Data • Given the need for data movement, how can we reliably do it well? 2/10/15 3
- 4. The Task of Large Scale Data Movement • Several different ways to look at a data movement task • People perspec<ve: – I am a member of a collabora<on – Our collabora<on has accounts with compute alloca<ons and data storage alloca<ons at a set of sites – I need to move data between those sites • Organiza<on/facility perspec<ve: – ANL, NCSA, NERSC, ORNL and SDSC are all used by the collabora<on – All these sites must have data transfer tools in common – I must learn what tools and capabili<es each site has, and apply those tools to my task • Note that the integra<on burden is on the scien<st! 2/10/15 4
- 5. Service Primi<ves • There is another way to look at data movement • All large-‐scale data movement tasks are composed of a set of primi<ves – Those primi<ves are common to most such workflows – If major sites can agree on a set of primi<ves, all large-‐scale data workflows will benefit • What are the common primi<ves? – Storage systems (filesystems, tape archives, etc.) – Data transfer applica<ons (Globus, others) – Workflow tools, if automa<on is used – Networks • Local networks • Wide area networks • What if these worked well together in the general case? • Compose them into common design paJerns 2/10/15 5
- 6. The Central Role of the Network • The very structure of modern science assumes science networks exist: high performance, feature rich, global scope • What is “The Network” anyway? – “The Network” is the set of devices and applica<ons involved in the use of a remote resource • This is not about supercomputer interconnects • This is about data flow from experiment to analysis, between facili<es, etc. – User interfaces for “The Network” – portal, data transfer tool, workflow engine – Therefore, servers and applica<ons must also be considered • What is important? Ordered list: 1. Correctness 2. Consistency 3. Performance © 2014, Energy Sciences Network 6 – ESnet Science Engagement ([email protected]) - 2/10/15
- 7. TCP – Ubiquitous and Fragile • Networks provide connec<vity between hosts – how do hosts see the network? – From an applica<on’s perspec<ve, the interface to “the other end” is a socket – Communica<on is between applica<ons – mostly over TCP • TCP – the fragile workhorse – TCP is (for very good reasons) <mid – packet loss is interpreted as conges<on – Packet loss in conjunc<on with latency is a performance killer – Like it or not, TCP is used for the vast majority of data transfer applica<ons (more than 95% of ESnet traffic is TCP) © 2014, Energy Sciences Network 7 – ESnet Science Engagement ([email protected]) - 2/10/15
- 8. A small amount of packet loss makes a huge difference in TCP performance Metro Area Local (LAN) Regional Con<nental Interna<onal Measured (TCP Reno) Measured (HTCP) Theoretical (TCP Reno) Measured (no loss) With loss, high performance beyond metro distances is essentially impossible © 2014, Energy Sciences Network 8 – ESnet Science Engagement ([email protected]) - 2/10/15
- 9. Design PaGern – The Science DMZ Model • Design paJerns are reusable solu<ons to design problems that recur in the real world – High performance data movement is a good fit for this – Science DMZ model • Science DMZ incorporates several things – Network enclave at or near site perimeter – Sane security controls • Good fit for high-‐performance applica<ons • Specific to Science DMZ services – Performance test and measurement – Dedicated systems for data transfer (Data Transfer Nodes) • High performance hosts • Good tools • Details at hJp://fasterdata.es.net/science-‐dmz/ 2/10/15 9
- 10. Context: Science DMZ Adop<on • DOE Na<onal Laboratories – Both large and small sites – HPC centers, LHC sites, experimental facili<es • NSF CC-‐NIE and CC*IIE programs leverage Science DMZ – $40M and coun<ng (third round awards coming soon, es<mate addi<onal $18M to $20M) – Significant investments across the US university complex, ~130 awards – Big shoutout to Kevin Thompson and the NSF – these programs are cri<cally important • Na<onal Ins<tutes of Health – 100G network infrastructure refresh • US Department of Agriculture – Agricultural Research Service is building a new science network based on the Science DMZ model – hJps://www.ro.gov/index?s=opportunity&mode=form&tab=core&id=a7f291f4216b5a24c1177a5684e1809b • Other US agencies looking at Science DMZ model – NASA – NOAA • Australian Research Data Storage Infrastructure (RDSI) – Science DMZs at major sites, connected by a high speed network – hJps://www.rdsi.edu.au/dashnet – hJps://www.rdsi.edu.au/dashnet-‐deployment-‐rdsi-‐nodes-‐begins • Other countries – Brazil – New Zealand – More 2/10/15 10
- 11. Context: Community Capabili<es • Many Science DMZs directly support science applica<ons – LHC (Run 2 is coming soon) – Experiment opera<on (Fusion, Light Sources, etc.) – Data transfer into/out of HPC facili<es • Many Science DMZs are SDN-‐ready – Openflow-‐capable gear – SDN research ongoing • High-‐performance components – High-‐speed WAN connec<vity – perfSONAR deployments – DTN deployments • Metcalfe’s Law of Network U<lity – Value propor<onal to the square of the number of DMZs? n log(n)? – Cyberinfrastructure value increases as we all upgrade 2/10/15 11
- 12. Strategic Impacts • What does this mean? – We are in the midst of a significant cyberinfrastructure upgrade – Enterprise networks need not be unduly perturbed J • Significantly enhanced capabili<es compared to 3 years ago – Terabyte-‐scale data movement is much easier – Petabyte-‐scale data movement possible outside the LHC experiments • 3.1Gbps = 1PB/month • (Try doing that through your enterprise firewall!) – Widely-‐deployed tools are much beJer (e.g. Globus) • Raised expecta<ons for network infrastructures – Scien<sts should be able to do beJer than residen<al broadband • Many more sites can now achieve good performance • Incumbent on science networks to meet the challenge – Remember the TCP loss characteris<cs – Use perfSONAR – Science experiments assume this stuff works – we can now meet their needs 2/10/15 12
- 13. High Performance Data Transfer -‐ Requirements • There is a set of things required for reliable high-‐performance data transfer – Long-‐haul networks • Well-‐provisioned • High-‐performance – Local networks • Well-‐provisioned • High-‐performance • Sane security – Local data systems • Dedicated to data transfer (else too much complexity) • High-‐performance access to storage – Good data transfer tools • Interoperable • High-‐performance – Ease of use • Usable by people • Usable by workflows • Interoperable across sites (remove integra<on burden) 2/10/15 13
- 14. Long-‐Haul Network Status • 100 Gigabit per second networks deployed globally – USA/DOE Na<onal Laboratories – ESnet – USA/.edu – Internet2 – Europe – GEANT – Many state and regional networks have or are deploying 100Gbps cores • What does this mean in terms of capability? – 1TB/hour requires less than 2.5Gbps (2.5% of 100Gbps network) – 1PB/week requires less than 15Gbps (15% of 100Gbps network) – hJp://fasterdata.es.net/home/requirements-‐and-‐expecta<ons – The long-‐haul capacity problem is now solved, to first order • Some networks are s<ll in the middle of upgrades • However, steady progress is being made 2/10/15 14
- 15. Local Network Status • Many ESnet sites now have 100G connec<ons to ESnet – 2x100G: BNL, CERN, FNAL – 1x100G: ANL, LANL, LBNL, NERSC, ORNL, SLAC • Capacity provisioning is much easier in a LAN environment • Security requires aJen<on (see Science DMZ) • Major DOE compu<ng facili<es have a lot of capacity deployed to their data systems – ANL: 60Gbps – NERSC: 80Gbps – ORNL: 20Gbps • Big win if sites use Science DMZ model 2/10/15 15
- 16. Progress So Far • There is a set of things required for reliable high-‐performance data transfer – Long-‐haul networks • Well-‐provisioned • High-‐performance – Local networks • Well-‐provisioned • High-‐performance • Sane security – Local data systems • Dedicated to data transfer (else too much complexity) • High-‐performance access to storage – Good data transfer tools • Interoperable • High-‐performance – Ease of use • Usable by people • Usable by workflows • Interoperable across sites (remove integra<on burden) 2/10/15 16
- 17. Local Data Systems • Science DMZ model calls these Data Transfer Nodes – Dedicated to high-‐performance data transfer tasks – Short, clean path to outside world • At HPC facili<es, they mount the global filesystem – Transfer data to the DTN – Data available on HPC resource • High-‐performance data transfer tools – Globus Transfer – Command-‐line globus-‐url-‐copy – BBCP • These are deployed now at many HPC facili<es – ANL, NERSC, ORNL – NCSA, SDSC 2/10/15 17
- 18. Data Transfer Tools • Interoperability is really important – Remember, scien<sts should not have to do the integra<on – HPC facili<es should agree on a common toolset – Today, that common toolset has a few members • Globus Transfer • SSH/SCP/Rsync (yes, I know – ick!) • Many niche tools • Globus appears to be the most full-‐featured – GUI, data integrity checks, fault recovery – Fire and forget – API for workflows • Globus is also widely deployed – ANL, NERSC, ORNL – NCSA, SDSC (all of XSEDE) – Many other loca<ons 2/10/15 18
- 19. More Progress • There is a set of things required for reliable high-‐performance data transfer – Long-‐haul networks • Well-‐provisioned • High-‐performance – Local networks • Well-‐provisioned • High-‐performance • Sane security – Local data systems • Dedicated to data transfer (else too much complexity) • High-‐performance access to storage – Good data transfer tools • Interoperable • High-‐performance – Ease of use • Usable by people • Usable by workflows • Interoperable across sites (remove integra<on burden) 2/10/15 19
- 20. Mission Scope and Science Support • Resource providers each have their own mission – ESnet: high-‐performance networking for science – ANL, NERSC, ORNL: HPC for DOE science users – NCSA, SDSC, et. al.: HPC for NSF users – Globus: full-‐featured, high-‐performance data transfer tools • No responsibility for individual science projects – Resource provider staff usually not part of science projects – Science projects have to do their own integra<on (see beginning of talk) • However, resource providers are typically responsive to user requests – If you have a problem, it’s their job to fix it – I propose we use this to get something done 2/10/15 20
- 21. Hypothe<cal: HPC Data Transfer Capability • This community has significant data transfer needs – I have worked with some of you in the past – Simula<ons, sky surveys, etc. – Expecta<on over <me that needs will increase • Improve data movement capability – ANL, NERSC, ORNL – NCSA, SDSC – This is an arbitrary list, based on my incomplete understanding – Should there be others? • Goal: per-‐Globus-‐job performance of 1PB/week – I don’t mean we have to transfer 1PB every week – But, if we need to, we should be able to – Remember, this only takes 15% of a 100G network path 2/10/15 21
- 22. What Would Be Required? • We would need several things: – Specific workflow (move dataset D of size S from A to Z, frequency F) – A commitment by resource providers to see it through • ESnet (+ other networks if needed) • Compu<ng facili<es • Globus • Is it 100% plug-‐and-‐play? No. – There are almost certainly some wrinkles – However, most of the hard part is done • Networks • Data transfer nodes • Tools • Let’s work together and make this happen! 2/10/15 22
- 23. Ques<ons For You • Would an effort like this be useful? (I think so) • Does this community need this capability? (I think so) • Are there obvious gaps? (probably, e.g. performance to tape) • Which sites would be involved? • Am I crazy? (I think not) 2/10/15 23
- 24. Thanks! Eli Dart Energy Sciences Network (ESnet) Lawrence Berkeley Na<onal Laboratory hJp://fasterdata.es.net/ hJp://my.es.net/ hJp://www.es.net/
- 25. Extra Slides 2/10/15 25
- 26. Support For Science Traffic • The Science DMZ is typically deployed to support science traffic – Typically large data transfers over long distances – In most cases, the data transfer applica<ons use TCP • The behavior of TCP is a legacy from the conges<on collapse of the Internet in the 1980s – Loss is interpreted as conges<on – TCP backs off to avoid conges<on à performance degrades – Performance hit related to the square of the packet loss rate • Addressing this problem is a dominant engineering considera<on for science networks – Lots of design effort – Lots of engineering <me – Lots of troubleshoo<ng effort 2/10/15 26
- 27. A small amount of packet loss makes a huge difference in TCP performance 2/10/15 Metro Area Local (LAN) Regional Con<nental Interna<onal Measured (TCP Reno) Measured (HTCP) Theoretical (TCP Reno) Measured (no loss) With loss, high performance beyond metro distances is essentially impossible