|
||||||||
|
||||||||
|
|
Công Cụ | Xếp Bài |
20-05-2015, 02:35 PM | #1 |
Guest
Trả Lời: n/a
|
For CEPH CLUSTER PLANNING
For CEPH CLUSTER PLANNING
What it is all about : If you think or discuss about Ceph , the most common question strike to your mind is "What Hardware Should I Select For My CEPH Storage Cluster ?" and yes if you really thought of this question in your mind , congratulations you seems to be serious about ceph technology and You should be because CEPH IS THE FUTURE OF STORAGE. Ceph runs on Commodity hardware , Ohh Yeah !! everyone now knows it . It is designed to build a multi-petabyte storage cluster while providing enterprise ready features. No single point of failure , scaling to exabytes , self managing and self healing ( saves operational cost ) , runs on commodity hardware ( no vendor locking , saves capital investment ) Ceph Overview :- The sole of ceph storage cluster is RADOS ( Reliable Autonomic Distributed Object Store ). Ceph uses powerful CRUSH ( Controlled Replication Under Scalable Hashing ) algorithm for optimize data placement , self managing and self healing. The RESTful interface is provided by Ceph Object Gateway (RGW) aks Rados GateWay and virtual disks are provisioned by Ceph Block Device (RBD) Ceph Overview - Image Credit : Inktank Ceph Components :- # Ceph OSD ( Object Storage Daemons ) storage data in objects , manages data replication , recovery , rebalancing and provides stage information to Ceph Monitor. Its recommended to user 1 OSD per physical disk. # Ceph MON ( Monitors ) maintains overall health of cluster by keeping cluster map state including Monitor map , OSD map , Placement Group ( PG ) map , and CRUSH map. Monitors receives state information from other components to maintain maps and circulate these maps to other Monitor and OSD nodes. # Ceph RGW ( Object Gateway / Rados Gateway ) RESTful API interface compatible with Amazon S3 , OpenStack Swift . # Ceph RBD ( Raw Block Device ) Provides Block Storage to VM / bare metal as well as regular clients , supports OpenStack and CloudStack . Includes Enterprise features like snapshot , thin provisioning , compression. # CephFS ( File System ) distributed POSIX NAS storage. Few Thumb Rules :-
Monitor Hardware Configuration :- Monitor maintains health of entire cluster , it contains PG logs and OSD logs . A minimum of three monitors nodes are recommended for a cluster quorum. Ceph monitor nodes are not resource hungry they can work well with fairly low cpu and memory. A 1U server with low cost processor E5-2603,16GB RAM and 1GbE network should be sufficient in most of the cases. If PG,Monitor and OSD logs are storage on local disk of monitor node , make sure you have sufficient amount of local storage so that it should not fill up. Unhealthy clusters require more storage for logs , can reach upto GB and even hundreds of GB if the cluster is left unhealthy for a very long time . If verbose output is set on monitor nodes, then these are bound to generate huge amount of logging information. Refer ceph documentation for monitor log setting. Its recommended to run monitor on distant nodes rather on all on all one node or on virtual machines on physical separated machines to prevent single point of failure. The Planning Stage :- Deploying a ceph cluster in production requires a little bit Homework , you should gather the below information so that you can design a better and more reliable and scalable ceph cluster to fit in your IT needs. These very specific to your needs and your IT environment. This information will help you to design your storage requirement better.
Collect as much information as possible during the planning stage , the will give all the answers required to construct a better ceph cluster. The Physical Node and clustering technique:- In addition to above collected information , also take into account the rack density and power budget , data center space pace cost to size the optimal node configuration. Ceph replicated data across multiple nodes in a storage cluster to provide data redundancy and higher availability. Its important to consider.
A general thumb of rule to calculate recovery time in a ceph cluster given 1 disk per OSD node is : Recovery Time in seconds = disk capacity in Gigabits / ( network speed *(nodes-1) ) # POC Environment -- Can have a minimum of 3 physical nodes with 10 OSD's each. This provides 66% cluster availability upon a physical node failure and 97% uptime upon an OSD failure. RGW and Monitor nodes can be put on OSD nodes but this may impact performance and not recommended for production. # Production Environment -- a minimum of 5 physically separated nodes and minimum of 100 OSD @ 4TB per OSD the cluster capacity is over 130TB and provides 80% uptime on physical node failure and 99% uptime on OSD failure. RGW and Monitors should be on separate nodes. Based on the outcome of planning phase and physical nodes and clustering stage you have a look on the hardware available in market as per your budget. OSD CPU selection :- |
|
|