|Model||IBM 3950 M2|
|Last updated||16 May 2019 10:47:11|
SCRAPPED AT Ujima House in Wembley. Borg was a stack of 4 linked machines that combine as one, resulting in half a terabyte of RAM and 64 Cores. It ran Linux designed for High Performance Computing on computationally expensive tasks.
- 1 Status
- 2 Naming
- 3 Access
- 4 Specs, Layout and Status
- 5 IP
- 6 RAS
- 7 ScaleXpandr
- 8 Logbook
- 9 Running Debian
- 10 Usage
- 11 Location
- 12 TODO
- 13 Potential uses
The BORG cluster was complete. The system was Ubuntu 12.04 LTS. We have 4 units in place (the maximum amount) with linux 12.04LTS installed with the usual hackspace credentials.
There are still some tasks to perform before the system is hackspace friendly. See the todo list below.
They're currently labeled as BORG1 through BORG6
Access is via ssh. This system requires physically plugging in the power cable and launching the complex through the RAS controller on Borg5
Steps to boot the cluster
- Firstly, locate the power cable. There is a blue, 16Amp socket on the wall. Nearby will be a power bar with a matching 16Amp plug. Attach this to the socket. (NEW: since now the plug is somehow seating very strongly, we unplug the primary cable from the power strip on the ground instead.) Borg should power up
- Each borg node has a dual powersupply, if all 8 PSU's are connected it may trip the breaker in the supply to the rack - if this happens the ups will start to beep. Switch the breaker back on if this happens!
- You can get away with only connecting one psu on each machine.
- The lights on borg should be FLASHING green. A slow flash on the power buttons. DO NOT press any of these buttons. All the machines need to be brought up via the RAS
- With a webbrowser on your laptop, attached to the hackspace network, navigate to http://172.31.24.170 (was "http://172.31.24.171, I believe this is the right IP address :S")
- This is Borg5, the headnode - you should see a webpage asking for username and password. Username USERID password PASSW0RD (note the zero in password as oppose to the letter 'o')
- Navigate to the last item on the left - RAS Partitions.
- You should see the partition on the right as a block diagram. All the nodes should be connected and in Blue. Find a button marked something like "boot partition". Press it.
- All the borgs should start whirring, more lights will come on and the power lights will turn green
- Wait about 10 minutes
Borg now has a static ip and gets username and password info from LDAP. It also mounts home directorys from colin. All hackspace members can sudo on Borg. You should have access to all the CPUs and memory in all 4 borg machines.
Specs, Layout and Status
IBM 3950 M2
- 16 cores total
- 4 x Xeon 7330: 4 cores @ 2.4 GHz "Tigerton" Intel Ark spec sheet
- 128 GB ram
The list reflects the layout in the rack.
- BORG5 - 4 CPUs, 128GB RAM - SN 99C5979 - 1.16 BIOS - HEAD Node
- BORG3 - 4 CPUs, 128GB RAM - SN 99C5980 - 1.16 BIOS
- BORG6 - 4 CPUs, 128GB RAM - 1.16 BIOS
- BORG4 - 4 CPUs, 128GB RAM - SN 99B3501 - 1.16 BIOS
BORG1 - 3 CPUs, 8GB RAM BORG2 - 3 CPUs, 8GB RAM
BORGs one and two are lower spec and cannot be linked to the cluster (max of 4 nodes) so should be cannibalised for spares and disposed of.
172.31.24.12 - RAS 172.31.24.11 - head node
The RAS II is a separate control system that is setup in the bios and accessible as soon as a BORG unit has power (i.e, if the green light is flashing OR solid). Using a web browser head to either
Username USERID password PASSW0RD
In order to link upto 4 borgs together we need to use the special cables. http://www.redbooks.ibm.com/redbooks/pdfs/sg247630.pdf - page 235 shows how this is done to create an SMP style set of nodes.
- Needs the non-free bnx2 firmware on a flash drive for the install to work, (could try to add it to the install initrd)
- Something something IBM Calgary IOMMU something something leads to DMA errors and the LSI MegaRaid raid card dosn't work, booting with "iommu=soft" makes it work but may not be ideal. search https://www.kernel.org/doc/Documentation/x86/x86_64/boot-options.txt for iommu
- The incantation seems to be: iommu=soft,calgary megaraid_sas.msix_disable=1
Upgrading the bios
Do a diskless boot, the go do Debian and then "Jessie amd64 Diskless for BORGs", log in as root (password 'root', this diskless setup is for testing only!), then:
cd ibm-bios/z/ ./lflash64
This is an upgrade to bios version 1.16
We also need to upgrade:
The RSA II thing: https://www-947.ibm.com/support/entry/portal/docdisplay?lndocid=MIGR-5086633 The FPGA (in the scaleXpander?) The BMC
We might be able to do it individually, or perhaps use the IBM UpdateXpress thing, which needs SUSE Linux Enterprise Server 11 x86-64 or Red Hat Enterprise Linux 6 x86-64.
We need to reset all the bios etc settings to there defaults.
good megaraid cli guide:
Testing with a normal, common-o-garden multimeter inline with the plug, on all 4 BORG units we have:
- 1.28amps at 240v - standby
- 7.8 amps peak on start
- 5 to 6 amps normal operation
- 9.48 amps with all 64 processors at 100%
Recent tests comparing Borg3 with AWS , rendering 30 Frames of a small Blender Scene with Blender 2.76 and maxing out the processors:
- $0.520 per On Demand Linux c1.xlarge Instance Hour at 12 hours
- Cost is roughly £4.60 at current exchange
- Spot instances are possible with Brenda or similar - at $0.07 the cost would be roughly 53p
- £0.12p per Kilowatt Hour
- With all 16 Cores operating at 100% Borg was drawing 560Watts
- To render the same scene took 206minutes
- Total Cost 23p
Borg (being a server) could add substantially to our electricity costs. To help this we can offset some of it by the retiring of other machines in the space. It has also been agreed that if it takes up more than 275W regularly then a decision about Borgs future will have to be made. If it goes over 300W we will have to re-think our strategy. These figures are a maximum and equate to about £27 per month in electricity costs. The power consumption will be regularly measured and anyone is welcome to report power consumption levels.
The Rack in the basement.
- Work out whats up with the 4th missing raid controller - looks like one node does not share it's pci devices?!?
- Play with the raid management thingy
- Work out what disks we have and where they are and what they do
- sdc and sdd are now free. (was a vmware install for emf)
- Re-install with debian?
Get it under ansible (with nfs home dirs and ldap users)done
- remove the FC cards we don't need. (or get a FC disk array!)
- channel bond some of the nics
- do something with the 10Gb Nic's?!?!?
- Rendering video and 3D
- Bio-informatics number crunching (bio-hackers?)
- realtime ray-tracing
- Radio FFT decoding in real-time (Cubesat related)