On Thu, 24 Jan 2019 10:49:14 +0300 Roman Simakov <roman.sima...@gmail.com> wrote:
> чт, 24 янв. 2019 г. в 10:25, liviuslivius > <liviusliv...@poczta.onet.pl>: > > > > Hi. > > > > I have thinked about oltp comparision, and i think that comparing > > e.g. classic vs superserver on RAM disc is wrong. Why? Because read > > from "disc" is as fast as read from cache. > > This case eliminates disc problems and allow to highlight other ones. > In other words RAM simulates the very good system IO like a PCIe-SSD. > Except not all SSDs are the same. There is an interesting video from linus tech tips here which gives an insight as to why: https://www.youtube.com/watch?v=OffzVc7ZB-o Basically most newer SSDs use more layers which means real world performance is now worse than for SLC based SATA SSDs from a few years back. The video reviews an intel ssd, but I've done (firebird-based) tests with a Samsung 970 EVO NVMe M.2 which also has poorer performance under heavy load than an older SATA based SSD. It is easy to think that m.2 must be better than PCIe which must be better than SATA but the reality is that the internal design of the NAND is the most important. It is interesting also to note that the load test I mentioned above with a RAM disk based DB also behaved differently to both the RAID10 HDD test and the tests with different SSDs. Number of CPUs did not seem to be as important. So, to summarize, in my opinion a RAM disc does simulate a very good system IO. But this is NOT synonymous with the IO from a real disc subsystem. Basically a RAM disc will reduce IO to the absolute minimum available with today's technology. But it is probably not the most suitable location to host a firebird database so the performance results would be invalid. Paul -- Paul Reeves http://www.ibphoenix.com Supporting users of Firebird Firebird-Devel mailing list, web interface at https://lists.sourceforge.net/lists/listinfo/firebird-devel