1 reply to this topic

[Olof Lagerkvist]

From the last couple of days discussions I got the idea of searching for RAMDISK benchmarks people may have done. I found this one, that shows a few interesting things.

http://fiehnlab.ucda...p;year:int=2011

Some observations:
* Generally, specialized RAMDISK drivers win over multi-purpose ones.
* For random access with large blocks, drivers allocating a large possibly partially pageable kernel memory block win over driver allocating a block scatter and distributes I/O requests among them.
* For random access with large blocks, drivers allocating a large immediately accessible, non-swappable, kernel memory block win over driver allocating a block scatter and distributes I/O requests among them.
* For random access with small blocks, drivers allocating a block scatter of guaranteed non-paged memory and distributes I/O requests among them wins over practically anything else.
* For sequential access drivers allocating contiguous memory by any technique win over any technique that need some kind of buffer swapping or remapping while accessing a long sequence of the disk.
* Read and write performance figures follow each-other quite well regardless of technique. This probably leads to the conclusion that it is in case of RAMDISKs not access to the memory itself that takes time, it is rather the memory remapping etc that the driver is required to do before accessing the memory that takes time.

It would also have been interesting to see how an ImDisk drive backed by the new awealloc driver had performed.

But what can be clearly seen here, there are probably far better alternatives than ImDisk in any case where a high-performance RAMDISK is required.

[MedEvil]

It would also have been interesting to see how an ImDisk drive backed by the new awealloc driver had performed.

Just use the named tools and make a test of the old driver and then with the new one. Afterward correct the results to fit with their test.