* Turn off IMU. Set _in_memory_undo to false. I've even seen Oracle support recommend this. It's a shame if you have to do this, but if it turns out there is a serious problem with the kernel code or how Oracle interacts with the operating system, you may need to completely disable IMU (and hope there is a patch).

* Increase CPU power. Increase the number of CPU cores, their speed, or available CPU cycles. The goal here is to give Oracle processes more CPU cycles. By "available CPU cycles," I mean to look at all operating system processes (which includes Oracle processes) and see if you can eliminate them, reduce their impact, or have them run during nonpeak times (that is, do some workload balancing).

* Increase the number of IMU latches. Keep in mind that just because the number of IMU latches increases, this does not mean Oracle will actually use them. Run a query like the one shown in Figure 7-18 to better understand the true usage situation. Here are a couple ways to increase the number of IMU latches: