When communication between parallel tasks occurs only over a single device such as css0 or en0, the application and the device are gated by each other. The device must wait for the application to fill a communication buffer before it can be transmitted and the application must wait for the device to transmit and empty the buffer before it can be refilled. Thus time is wasted by each waiting for the other. The technique of striping refers to using two communication paths to implement a single communication path as perceived by the application. As the application sends data, it fills up a buffer on one device. As that buffer is transmitted over the first device, the application's data begins filling up the second buffer and the application perceives no delay in being able to write. When the second buffer is full, it begins transmission over the second device and the application reverts to the buffer of the first device. Much, if not all of the buffer on the first device has been transmitted while the application wrote to the second buffer so the application waits for a minimal amount of time or possibly does not wait at all. LoadLeveler supports job requests to use striped communication with the virtual device csss. When one or more switch adapters are specified on a node, a csss adapter is automatically created. This csss adapter is requested on the job's network statement when striping is desired.
Jobs that request striping through the csss device (both user space and IP) can be submitted to nodes with only one switch adapter. In that situation, the result is the same as if the job requested css0.
| Note: | When configured, a multi-link address is associated with the virtual ml0 device. The IP address of this device is the multi-link address. The llextSDR program will create a stanza for the ml0 device that will appear similar to Ethernet or token ring adapter stanzas except that it will include the multilink_list keyword that lists the adapters it performs striping over. As with any other device with an IP address, the ml0 device can be requested in IP mode on the network statement. Doing so would yield a comparable effect to requesting csss IP except that no checking would be performed by LoadLeveler to ensure the associated adapters are actually working. Thus it would be possible to dispatch a job that requested communication over ml0 only to have the job fail because the switch adapters that ml0 stripes over were down. |