Messages in this thread | | | Date | Fri, 08 Apr 2005 00:49:10 -0700 | From | "H. Peter Anvin" <> | Subject | Re: Kernel SCM saga.. |
| |
Daniel Phillips wrote: > On Friday 08 April 2005 03:05, Rogan Dawes wrote: > >>Take a look at >>http://www.linuxshowcase.org/2001/full_papers/ezolt/ezolt_html/ >> >>Abstract >> >>GNU libc's default setting for malloc can cause a significant >>performance penalty for applications that use it extensively, such as >>Compaq's high performance extended math library, CXML. The default >>malloc tuning can cause a significant number of minor page faults, and >>result in application performance of only half of the true potential. > > > This does not smell like an n*2 suckage, more like n^something suckage. > Finding the elephant under the rug should not be hard. Profile? >
Lack of hysteresis can do that, with large swats of memory constantly being claimed and returned to the system. One way to implement hysteresis would be based on a decaying peak-based threshold; unfortunately for optimal performance that requires the C runtime to have a notion of time, and in extreme cases even be able to do asynchronous deallocation, but in reality one can probably assume that the rate of malloc/free is roughly constant over time.
-hpa - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| |