When the throughput of the SBRC system might temporarily be greater than the throughput of the downstream, and you want reliable transport so that the accounting gets downstream eventually, you can use the spooled accounting mechanism. In this efficient method, the packets are sent downstream synchronously about as fast as the downstream can go plus latency, but are sent outside of the transaction path. The overhead is in additional I/O to the local file system, which is approximately double the same throughput used for the equivalent network operation. For a 2000 RADIUS transaction, 2000 is written to the file, then later the file is read and the 2000 is written to the network. Currently, spooled accounting logs are drained by one thread per realm and have one in-flight transaction outstanding. Contact the Juniper Network Sales Engineer or manager for additional throughput of multiple in-flight transactions or multiple threads per realm.
If the average performance requirements of the system are greater than the downstream realm can handle, the backlog can be significant, and delays in draining the local spool to the downstream proxy can reach a point where the downstream will never catch up. See the SBR Carrier Reference Guide, [SpooledAccounting] section, for more information.