NodesAllocation
Existing machines
New machine node matrix
Node Matrix:
Notes:
- (1) Public IP is needed if we want to allow outside users direct access
How to create a KVM host with cobbler:
http://twiki.mwt2.org/bin/view/Admins/NewCobblerNode
New machines 1 (R610 with 2 MD1200)
The available disk on the MD1200 boxes is allocated in 3 equal parts (dCache, Hadoop, xrootd).
The R610 will use one drive to run the host OS (SL 5.5) and the other 5 drives (RAID configuration - LVS) will be used for the virtual machines. There are 9 VMs, 3 each for dCache, Hadoop, xrootd
Storage (virtual) nodes include:
- 3 VMs for dcache pools
- 3 VMs for xrootd pools
- 3 VMs for hdfs pools
New machines 2 (8 R610)
These R610 servers will be installed with:
- VMware ESXi on itb-esx3 (installation CD image available from vmware.com)
- SL 5.5 and KVM in itb-kvmX: one drive to run the host OS (SL 5.5) and the other 5 drives (RAID configuration - LVS) will be used for the virtual machines
- SL 5.5 like Tier 2/3 worker nodes on itb-cXXX: probably one drive to run the OS (SL 5.5) and the other 5 drives (RAID configuration - LVS) for the scratch space
- for itb-dyn check what is needed for dynamic provisioning (create/destroy VM dynamically); check Nimbus requirements (http://www.nimbusproject.org/ )
Headnode services for ITB and the virtual tier 3 include:
- PBS scheduler
- Condor headnode (collector/negotiator)
- GUMS
- ITB CE Headnode
- VTB CE Headnode
- Tier 3 CE
- Tier 3 SE (server varies)
- Tier 3 GUMS
- Tier 3 UI
- VO frontend (for Glidein-WMS job submission)
- Engagement submit host (with OSG MM submission)
- Campus router
We have currently 12 services in the list above. We may consider allocating one of the 2 unallocated R610 to host headnode services as well (itb-kvm6.uchicago.edu, with int/ext cabling).
We discussed having one management/service node (possibly one of the previous ITB nodes) to use as:
- login node
- NFSv4 server for home directories
- Ganglia monitor
Services "borrowed" from the Tier 2/Tier 3 infrastructure:
- Cobbler
- OSG_APP with ATLAS releases
IP Reallocation
IPs from the grid school machines have been reallocated as follows:
Testing
Interesting tests that we discussed.
IO testing
- IO comparison:
- local IO (on host OS) vs
- direct IO from VM (aka raw disks, access to the HW partition) vs
- virtual disks on VM
Worker nodes
- Compare real WN with virtual WN
Dynamic Provisioning
Dynamic allocation tests (
http://www.nimbusproject.org/)
Virtual machine infrastructure
Compare different virtualization software frameworks, i.e. ESXi vs KVM.
--
MarcoMambelli - 20 Jan 2011