Versions Compared

Old Version 106

changes.mady.by.user K Scott Rowe

Saved on

compared with

New Version Current

changes.mady.by.user K Scott Rowe

Saved on

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

Table of Contents

...

Budget of $330,000

To Do

  • Talk with SSA and VLASS about how we actually use these remote clusters (e.g.Staging, flocking, etc)
  • Buy test system
  • Talk with Pueto Rico about their technical situation
  • Think about getting drives from third party instead of Dell and sparing them ourselves.
  • Think about getting the same drives for the OS that we get for the data array.  It's simpler but perhaps slower.
  • Who is going to put all the RAM and drives in the servers?
  • Work on a price spreadsheet
  • Learn Ansible?

Timeline

  • Buy test system as soon as practical
  • Buy production system about 6 months after test system
  • Receive production system about 7 months after test system
  • Install production system about 10 months after test system
  • Running about 12 months after test system


Data Path

This is conceptual at this point.  Need to talk with SSA and VLASS about this.

  • We pre-stage data on the remote head node
  • We then either submit a job locally and it flocks to the remote site or we login to the remote site and submit from there.
    • Can we use a nifty filesystem to simplify this (Ceph or that LHC fs)?
    • This might be a good phase2 problem to solve.
    • Is this kinda what nraorsync does?
  • The remote execute hosts transfer data from the remote head node
  • The job uploads resulting data to the remote head node
  • We retrieve data from the remote head node


Using

  • Get NRAO jobs on the remote racks.  This may depend on how we want to use these remote racks. If we want them to do specific types of jobs then ClassAd options may be the solution. If we want them as overflow for jobs run at NRAO then flocking may be the solution. Perhaps we want both flocking and ClassAd options.  Actually flocking may be the best method because I think it doesn't require the execute nodes to have external network access.
    • Staging and submitting remotely?
    • Flocking?
    • Classad options?  I think this will require the execute hosts to have routable IPs because our submit host will talk directly to them and vice-versa.  Could CCB help here?
    • Other?
  • Remote HTCondor concerns
    • Do we want our jobs to run an NRAO user like vlapipe, or nobody?
    • Do we want remote institution jobs to run as the remote institution user, some dedicated user, or nobody?
  • Need to support 50% workload for NRAO and 50% workload for remote institution.  How?
  • Share disk space on head node 50% NRAO and 50% remote institution.  It might be nice to use something like quotas or LVM so that we can change the disk usage dynamically for times when either NRAO or local needs more space.


Documentation

  • A projectbook like we did for USNO could be appropriate
  • Process diagrams (how systems boot, how jobs get started from NRAO and run, how remote institutions start jobs, etc)


Networking

HTCondor flocking requires

  • From local schedd to remote collectord on condor port 9618
  • From remote negotiator and execute hosts to local schedd.  Here the execute hosts can be NATed.
  • From local shadow to remote starterd.  Use CCB.  It allows execute hosts to live behind firewall and be NATed.

Non-flocking just requires ssh access from probably mcilroy, and to gibson.

NRAO side

  • NRAO -> remote head node on port 22 (ssh)
  • Submit Host -> remote head node (condor_collector) on port 9618 (HTCondor) for flocking
  • Submit Host <- remote head node (condor_negotiator) on port 9618 (HTCondor) for flocking
    • mcilroy has external IPs (146.88.1.66 for 1Gb/s and 146.88.10.66 for 10Gb/s).  Is the container listening?
  • Submit Host <- remote execute hosts (condor_starter) on port 9618 (HTCondor) for flocking
  • Submit Host (condor_shadow) -> remote execute hosts (condor_starter) on port 9618 (HTCondor) for flocking.  CCB might alleviate this.

Remote side

  • Head node <- from nrao.edu on port 22 (ssh)
  • Head node -> revere.aoc.nrao.edu on port 25 (smtp)
  • Head node -> NRAO Submit Host on port 9618 (HTCondor) for flocking
  • Head node <- NRAO Submit Host on port 9618 (HTCondor) for flocking
  • Execute node -> NRAO Submit Host on port 9618 (HTCondor) for flocking.  Execute host may be NATed.
  • Execute node -> gibson.aoc.nrao.edu on port 22 (ssh) for flocking with nraorsync.  Execute host can be NATed.


Services

  • DNS
    • What DNS domain will these hosts be in?  nrao.edu? remote-institution.site? other?
    • Will this vary depending on site?
    • 2022-10-26 krowe: it is looking like the institution will own the equpment.  Either they buy it with their own money like UPR-M or AUI gives them a grant and they buy it.  Either way, they own it.  So, I think we can expect the hosts to be in their DNS domain.  Which is probably for the best.  We can make CNAMEs for each head node if needed.
    • So what IP range should we use?  That may depend on the site as each site may use non-routable IP ranges differently.
  • DHCP
  • SMTP
  • NTP or chrony
    • What timezone should these be in?  I think the choices are
      • Mountain - Perhaps the most convenient for NRAO users and consistant between the sites.
      • Local - Makes the most sence to the local users but means differences between the sites.
      • UTC - equally annoying for all.
  • NFS
  • Directory Server
    • NIS?  Probably not.  RHEL9 will not support NIS.
    • OpenLDAP
    • 389 Directory Server? (previously Fedora Directory Server)
    • Identity Management
    • FreeIPA
    • How do we handle accounts?  I think we will want accounts on at least the head node.  The execution nodes could run everything as nobody or as real users.  If we want real users on the execute hosts then we should use a directory service which should probably be LDAP.  No sense in teaching folks how to use NIS anymore.
      • remote institution accounts only?
    • 2022-10-26 krowe: RHEL8 and later don't come with OpenLDAP anymore.  Red Hat wants you to use either their 389DS or IDM or RHDS or some other thing that gets them money.  It's all very confusing
  • ssh
  • rsync (nraorsync_plugin.py)?
  • NAT so the nodes can download/upload data?
  • TFTP (for OSes and switch)
  • condor (port 9618) https://htcondor-wiki.cs.wisc.edu/index.cgi/wiki?p=HowToMixFirewallsAndHtCondor
  • nagios
  •  ganglia
    • Ganglia hasn't been updated since 2015 so perhaps it is time to look for something else.
    • Prometheus/Graphana
    • Zabbix


Operating System

  • Must support CASA
  • Will need a patching/updating mechanism
  • How to boot diskless OS images
  • What Linux distrobution to use?
    • Can we use Red Hat with our current license?  I have looked in JDE and I can't find a recent subscription.  Need to ask David.
      • We have a 1,000 FTE license with up to 20,000 installations allowed.  But since we are either selling the equipment to the institution or asking them to buy it themselves, at least with UPRM, I don't see how we can use our RHEL license legally.  Asking each institution to aquire an RHEL license sounds like a recepie for failure so I think open source OS is the answer.
    • Should we buy Red Hat licenses like we did for USNO?
      • USNO is between $10K and $15K per year for 81 licensed nodes.  This may not be an EDU license.
      • NRAO used to have a 1,000 host license for Red Hat but I don't know what they have now.
      • I don't want to maintain licenses for up to 10 differenct installs.  I don't think the institutions will want to purchase and maintain a license.
    • Do we even want to use Red Hat?
      • Alternatives would be Rocky Linux or AlmaLinux or CentoOS Stream
    • Some sites will own their equipment like UPR-M.  Probably most sites the equipment will be owned by NRAO.
  • What version do we use RHEL7 or RHEL8 or RHEL9?  Remember CASA needs to support it.
  • What OSes is CASA is verified against?  I am pretty sure RHEL but what about CentOS or Rocky or ALMA, etc?.
  • What version of CASA does VLASS need?
  • The cost of RHEL is pretty small compared to the hardware.
  • UPRM has their own money but the other institutions will either get a grant or money from us so we can say what OS they use and pay for.
  • Should pull in Matthew or Schlake on this decision.
  • Should we use Ansible for deployments?
  • 2022-10-21 krowe: jkern talked to business office and they prefer that AUI gives the money to Morgan State and Morgan State buys the equipment.  So Morgan State will own it.  NRAO can stipulate you will only get the money if you buy what we recommend.
  • 2022-10-24 krowe: CASA is verified against RHEL8. CentOS Stream 8 is a constantly moving target.  There is no CentOS Stream 8.1 or 8.2 it is always the cutting edge of RHEL.  I don't like that.  I would much rather have version numbers that I can compare to RHEL.  So that is a vote against CentOS Stream and for Rocky or Alma.  I just checked Scientific Linux (maintinaed by Fermilab and CERN) and they are moving to CentOS Stream8.  So there will not be a Scientific Linux 8.

Third party software for VLASS

  • CASA what version?
  • HTCondor
  • Will need a way to maintain the software
    • stow, rpm, modules, containers?

Third party software for remote institution

  • Will need a way to maintain software for the remote institution site
  • Will need a way to maintain the software
    • stow, rpm, modules, containers?


Management Access

  • PDU
  • UPS
  • BMC/IPMI
  • switch

Maintenance

  • replace disk (remote institution admin)
  • replace/reseat DIMM (remote institution admin)
  • replace power supply (remote institution admin)
  • NRAO may handle replacement hardware. Drop ship. Spare ourselves?
  • Patching OS images (NRAO)
  • Patching third party software like CASA and HTCondor (NRAO)
  • Altering OS images (NRAO)

Hardware

Operating System

  • Must support CASA
  • Will need a patching/updating mechanism
  • How to boot diskless OS images
  • What Linux distrobution to use?
    • Can we use Red Hat with our current license?  I have looked in JDE and I can't find a recent subscription.  Need to ask David.
    • Should we buy Red Hat licenses like we did for USNO?
      • USNO is between $10K and $15K per year for 81 licensed nodes.  This may not be an EDU license.
      • NRAO used to have a 1,000 host license for Red Hat but I don't know what they have now.
    • Do we even want to use Red Hat?
      • Alternatives would be Rocky Linux or AlmaLinux since CentOS is essentially dead

Third party software for VLASS

  • CASA
  • HTCondor
  • Will need a way to maintain the software
    • stow, rpm, modules, containers?

Third party software for Local

  • Will need a way to maintain software for the local site

Services

  • DNS
  • DHCP
  • SMTP
  • NTP
  • NFS?
  • LDAP?  How do we handle accounts?  I think we will want accounts on at least the head node.  The execution nodes could run everything as nobody or as real users.  If we want real users on the execute hosts then we should use a directory service which should probably be LDAP.  No sense in teaching folks how to use NIS anymore.
    • Local accounts only?
  • ssh
  • rsync (nraorsync_plugin.py)
  • NAT so the nodes can download/upload data
  • TFTP
  • condor (port 9618) https://htcondor-wiki.cs.wisc.edu/index.cgi/wiki?p=HowToMixFirewallsAndHtCondor
  • ganglia
  • nagios

Management Access

  • PDU
  • UPS
  • BMC/IPMI
  • switch

Maintenance

  • replace disk (local admin)
  • replace/reseat DIMM (local admin)
  • replace power supply (local admin)
  • NRAO may handle replacement hardware. Drop ship. Spare ourselves?
  • Patching OS images (NRAO)
  • Patching third party software like CASA and HTCondor (NRAO)
  • Altering OS images (NRAO)

Hardware

  • Cabinet Rack:  Doors front and rear locking with mesh.  Width: 19".  Height: 42U is most common.  Depth: 42" is most common.  Rack must support at least 2,000 lbs static load
  • PDU: one PDU or two PDUs? What plug? What voltage? This may very across sites.  What if the site has two power sources?
  • UPS: for just the head node and switch?  This may depend on the voltage of the PDUs.
  • UPS: for just the head node and switch?  This may depend on the voltage of the PDUs.
  • Switch: 10Gb/s.
  • Environmental Monitoring:  Could the PDU do this?
  • Switch: Cisco Catalyst C9300X-48TX

    • 48 Data, 48x 10G Multigigabit

    • 100M, 1G, 2.5G, 5G, or 10 Gbps
    • Switching capacity 2,000 Gbps
    • Nine optional Modular Uplinks 100G/40G/25G/10G/1G
    • Redundant Power Supply 715 W
    • $12K
  • Environmental Monitoring:  Add-on to the APC PDU
  • KVM: rackmount, not remote, and patch cablesKVM: rackmount, not remote, and patch cables
  • Ethernet cables:
  • Power cables: single or Y calbes cables depending on number and types of PDUs and two number of power sources.
  • Head Node: lots of disk .  Do the locals have access to this disk space?  Maybe not.and RAID
    • iDRAC: Ask CIS what they recommend
    • Memory: at least 32GB of RAM to help cache the OS image.  64GB would be even better.
    • Storage mdRAID, ZFS, Btrfs, RAID card?  Do we want both boot and data arrays to be the same type?
      • OS/OSimages, RAID1 +1 with or without spare? (3 disks), about 1TB
        • OS: We have been making 40GB partitions for / for over a decade and that looks to still work with RHEL8.
        • Swap: 0 or 8GB at most
        • /export/home/<hostnmae>: services and diskless_images
      • Working data/software (local remote institution and nrao), RAID6 +1 w/spare or RAID7 (9 disks), about 72TB
        • An SE imaging input data size is about 10GB per job
        • We need maybe 20TB+ of total space or more so maybe 60TB/2
        • Carve into two partitions (NRAO data and NRAO software, Local remote institution data and Local remote institution software) each partition has data and software directories.
    • Networking: May need more than one port.  One for internal networking to nodes and one for external Internet access.
  • 30 1U nodes or 15 2U nodes or mix?  NVMe drives for nodes.  Swap drive?
    • Dual ~3GHz, ~12Core CPUs
    • 512GB or more RAM
    • NVMe for scratch and swap 6TB or more
    • Draws about 400W max
  • GPUs:  Do we get GPUs?  Do we get 1U nodes with room for 1 or 2 Tesla T4 GPUs or 2U nodes with room for 1 or 2 regular GPU?

      Using

      • Get NRAO jobs on the remote racks.  This may depend on how we want to use these remote racks. If we want them to do specific types of jobs then ClassAd options may be the solution. If we want them as overflow for jobs run at NRAO then flocking may be the solution. Perhaps we want both flocking and ClassAd options.  Actually flocking may be the best method because I think it doesn't require the execute nodes to have external network access.
        • Flocking?  What are the networking requirements?
        • Classad options?  I think this will require the execute hosts to have routable IPs because our submit host will talk directly to them and vice-versa.  Could CCB help here?
        • Other?
      • Remote HTCondor concerns
        • Do we want our jobs to run a an NRAO user like vlapipe or nobody?
        • Do we want local jobs to run as the local user, some dedicated user, or nobody?Remote HTCondor concerns
      • Need to support 50% workload for NRAP and 50% workload for local.  How?
      • Share disk space on head node 50% NRAO and 50% local
        • Two partitions: one for NRAO and one for local?

      Documentation

      • A projectbook like we did for USNO could be appropriate
      • Process diagrams (how systems boot, how jobs get started from NRAO and run, how locals start jobs, etc)

      Shipping

      • Cooling:  Assuming a 30-node cluster, we will need about 12kW or 3.5 tons or 42,000 BTU/h of cooling.


      Shipping

      • Drop ship everything to the site and assemble on site.  This will require an NRAO person on site to assemble with a pre-built OS disk for the head node.  I think this is too much work to do on site.
        • Install DIMMs in nodes
        • Install NVMe drives in nodes
        • Rack everything
        • Cable everything
        • Configure switch
        • Install/Configure OS
        Drop ship everything to the site and assemble on site
      • Ship everything here and assemble then ship a rack-on-pallet.
      • Mix the two. Ship minimal stuff here (head node, switch, couple of compute nodes, etc) and configure and drop ship most of the nodes to the site.
        • Re-ship head node, switch, compute nodes to site
        • Re-ship memory and drives to site
      • A person from the remote site could travel to NM or CV to see the test system and get instruction.


      Other

      • Keep each rack pod as similar to the other racks pods as possible.
      • Test system at NRAO should be one of everything.
      • Since we are making our own little OSG, should we try to leverage OSG for this or not?  Or do we want to make each POD a pool and flock?
      • Should we try to buy as much as we can from one vendor like Dell to simplify things?
      • APC sells a packaged rack on a pallet ready for shipping.  We could fill this with gear and ship it.  Not sure if that is a good idea or not.  We will not be able to move the unit into the server room while still on the pallet because no doorway is tall enough.  We would have to roll it off the pallet (it comes with a ramp and the rack is on casters) move it into the server room, fill and configure it, roll it out of the server room, roll it back onto the pallet, probably remove the bottom server(s) so we can attach it to the pallet, then re-add the bottom server(s).  We could use the double glass doors for this but there is a lip on the transition.  We could use the doors in the PRA closet as it has no lip but would require a lot of moving of shelves and stuff.
      • APC NetShelter SX packaged:
        • On Pallet: Height 85.79in (2179mm) Width 43.5in (1105mm)
        • On Casters: Height 78.39in 1991mm) Width 23.62in (600mm)
      • NRAO Dimensions
        • Double Glass doors: Height: 80in (2032mm) (because of the 2in maglock)
        • NRAO-NM wide server doors: Height: 83in (2133mm) Width: 48in (1187mm)
      • I could start prototyping now using AWS.
      • If jobs are submitted from the remote head node, does that mean SSA will want a container on that remote head node?



      Site Questions

      • Voltage in server room (120V or 208V or 240V)
      • Receptacles in server room (L5-30R or L21-30R or ...)
      • Single or dual power feeds?or dual power feeds?
      • Is power from below or from above?
      • How stable is their power?
      • Is there a UPS?
      • Is there a generatorIs power from below or from above?
      • Door width and height and path to server room.
        • Can a rack-on-pallet fit upright?  Height: 85.79inches (2179mm) Width: 43.5inches (1105mm)
        • Can a rack-on-casters fit upright?  Height: 78.39inches (1991mm) Width: 23.62inches (600mm)
        • NRAO-NM wide server door Height: 84inches (2108mm) Width: 46.75inches (1219mm)
      • Firewalls
      • How are you going to use this?
      • Do you care if this is in your DNS zone or ours?
      • Is NAT available for the execute hosts?


      Resources

      • USNO correlator (Mark Wainright)
      • VLBA Control Computers (William Colburn)
      • Red Hat maintenance (William Colburn)
      • Virtual kickstart (William Colburn)
      • Switch models and ethernet (Jeff Long)
      • HTCondor best practices (Greg Thain)
      • OSG (Lauren Michael)
      • SDSC at UCSD
      • TACC at UT Austin
      • IDIA https://www.idia.ac.za/

      Timeline

      • June buy
      • Receive by July
      • Pay by Aug
      • Install by Dec
      • Running in Jan. 2023