One of the possible ways to speed up scientific research projects which JINR and its Member State organizations participate in is to join computational resources. It can be done in a few ways one of which is to build distributed cloud infrastructures integrating local private clouds of JINR and organizations from its Member States. To implement such scenario, a cloud bursting based approach was chosen since it provides more flexibility in resources management of each cloud integrated into such distributed infrastructure. For the time being a few private clouds from some organizations from the JINR Member States have been already integrated with the JINR one with help of custom cloud bursting driver developed by the JINR team. Various aspects of that activity are covered in more details in the article as well as implemented and planned changes in the JINR cloud such as a smart scheduler deployment in production, migration from DRBD high-availability (HA) to native OpenNebula HA, Ceph storage deployment and VMs' disks migration to it and others.
Kazakhstan
Astana, Kazakhstan
Yerevan,
Armenia Tbilisi, Georgia
10 10 256 20 20
To achieve scientific results faster, the research organizations join own resources. There are
few ways to do that. One of them is to join one of the worldwide grid infrastructure (e.g. WLCG).
However, the grid site deployment and maintenance can be difficult for small research institutions in
contrast to a local private cloud. For that purpose a custom driver [
At the moment of writing the article a progress with clouds integration of each of the listed in the Table 1 organizations is at the various stage. facility (see Figure 2). Possible candidates for workload management system (WMS) of such a glue layer can be PanDA, DIRAC WMS, Vac/VCycle, VPManager/VCondor. 2. Upcoming high availability setup based on a Raft consensus algorithm
The JINR cloud [
Apart from that a Ceph-based [
All these at the end should lead to the new HA OpenNebula JINR cloud architecture shown on the Figure 3. There are the following types of the OpenNebula 5.4 testbed servers shown on the Figure 3 Cloud worker nodes (CWNs) which host virtual machines (VMs) and containers (CTs) are marked on the Figure 3 by numeral “1” in a grey square; Cloud front-end nodes (CFNs) where all core OpenNebula services including database, scheduler and some other ones are deployed (such hosts are marked on the Figure 3 by black numeral “2” inside the same color square); Cloud storage nodes (CSNs) based on Ceph SDS for keeping VMs’ and CTs’ images as well as users’ data (marked by black numeral “3” inside the same color square).
All these servers are connected to the same set of networks: JINR public and private subnets (they are marked on the Figure 3 by blue lines near which the numeral “1” is placed in the same color circle and signed as “JINR pub and priv subnets”); An isolated private network dedicated for a SDS traffic (dark green lines with numeral “2” in a circle of the same color and signed as “Priv subnet for storage”; Management network (black lines with numeral “3” in a circle of the same color singed as “management subnet”).
All network switches excluding those which are for the management network have 48 ports with 10 GbE each as well as four 40 Gbps SPF-ports for uplinks.
All CSNs apart from HDDs disks for data have SSD disks for caching.
In addition to increasing a common availability and reliability of the JINR cloud operation it is planned that the JINR cloud in such a new configuration will sufficiently improve users’ experience in terms of VMs/CTs deployment and migration times due to hosting its images on the fast SDS shared across CWNs as well as by providing users with the additional disk space to keep their data on the same SDS permanently and even more reliably.
The JINR cloud is actively developing: the new HA setup based on Raft consensus algorithm and Ceph-based SDS are under evaluation, new hardware is bought and put into production as well as private clouds of partner organizations from the JINR Member States are integrated into the united distributed cloud infrastructure. A research work on possible alternatives for clouds interconnection is in progress. More experiments show interest in the JINR cloud resources usage. Their needs and workflows are under evaluation to implement in our cloud environment.