Databases Reference
In-Depth Information
JOB_ID. The global join index may reduce the amount of cross-node traffic even when
collocation is not perfect. This index type will require additional storage for the RIDs
that point back to the base table. Generally speaking, the storage requirements for the
RIDs are a relatively small overhead when the index keys are more than 8 bytes wide.
The wider the index key, the less concern the designer needs to have over the overhead
of the RIDs in the global join index.
6.7.3 The Internode Interconnect
There will always be some degree of communication between shared-nothing partitions.
Even when data is perfectly collocated, results from each data node will need to be
merged into a result set to be returned to the calling application. Three major points are
worth mentioning in this respect:
•
Fast interconnects, such as Fibre Chanel or InfiniBand.
•
Te r a d a t a B y N E T.
•
Logical partitioning.
The first and most obvious strategy is to make sure your MPP database uses a
high-performance interconnect such as a high-speed Fibre Channel interconnect. The
four major generic interconnect alternatives are Fibre Channel, InfiniBand, Gigabit
Ethernet, and ATM. Table 6.4 reproduces a summary comparing the technologies
courtesy of the Fibre Channel Industry Association (www.fibrechannel.org). Fibre
channel has gained adoption in recent years for applications that require high-band-
width, reliable solutions that scale from small to very large.
Teradata AMPs use a specialized interconnect called the ByNET that has properties
that are well suited to data warehousing. The ByNET can create dynamic groups, mes-
saging, and sorting directly within the interconnection network. The following is a
paraphrased tech note from the Teradata's web pages that summarizes the capabilities
and advantages of the ByNET.
The BYNET works like a phone switch, quite different from the typical network. Its
switched “folded banyan” architecture delivers additional network bandwidth for each
node added to the configuration. Each connection to a node delivers 120 MB/sec. A 2-
node system has a 240 MB/sec. interconnect; 4 nodes, 480 MB/sec.; 8 nodes, 960 MB/
sec.; 100 nodes, 12 GB/sec. The current BYNET implementation can support up to 512
nodes (1 to 4 PB) but the design is capable of 4,096 nodes (8 to 32 PB) if a customer is
ready to solve a problem of that magnitude.
In addition to delivering fully scalable point-to-point bandwidth, the BYNET imple-
ments specialized functions not offered by other networking solutions. It can broadcast-
deliver a single message to some or all of the nodes in the MPP configuration. There are
many database functions that need to be performed on all nodes at once. With broadcast,
