SSD Caching

sanity · Thu Dec 27, 2012 8:41 am

Hi,

I've been testing the free version of StarWind in the last couple of weeks with a view to running a 3 node HA cluster next year.

I've a few questions if I may?

I've been using the host on an X79 motherboard connecting to a Revodrive 3 X2 PCIe SSD with 10Gb ethernet. I've found that when testing using sqlio or iometer that if I host a virtual disk on the SSD it performs much faster with no RAM caching than it does with RAM caching, from what I remember it's about 30% slowing with RAM caching. Is it best to use no cache with SSDs?

How much difference would running with 64GB of Ram compared to 32GB Ram? Currently running 32GB, but assume that's only really effective if caching a slower harddrive? It's a shame that in order to use more than 32GB you need a bigger version of windows rather than just Standard edition.... Out of interest why can we not use StarWind with win 7?

With regards to CPU, would a faster CPU be better than one with more cores? I can never seem to max out the CPU despite the network and SSD not being maxed out. even with caching turned off benchmarking is about 70% performance of native. any thoughts?

Kind regards..

sanity · Thu Dec 27, 2012 11:25 am

Ok, so turns out you can install the service on Windows 7.

What would be the advantage of installing on windows server 2008 r2 instead of windows 7? Win 7 Pro supports up to 192GB ram and costs about £100 whilst 2008r2 costs more than £800 to support more than 32GB ram...

thanks.

Thu Dec 27, 2012 9:30 pm

It's very difficult to measure cache pros and cons with pure synchetic tests. For writes you do saturate cache quite soon and it works only adding latency by purging old dirty cache lines. For read you need to have data in cache and synchetic tests just read random content making cache useless (it does not work as a cache). So I'd recommend to test with a real workload.

More RAM you throw in - more cache we'll allocate.

CPU is never a bottleneck these days.

sanity wrote:Hi,

I've been testing the free version of StarWind in the last couple of weeks with a view to running a 3 node HA cluster next year.

I've a few questions if I may?

I've been using the host on an X79 motherboard connecting to a Revodrive 3 X2 PCIe SSD with 10Gb ethernet. I've found that when testing using sqlio or iometer that if I host a virtual disk on the SSD it performs much faster with no RAM caching than it does with RAM caching, from what I remember it's about 30% slowing with RAM caching. Is it best to use no cache with SSDs?

How much difference would running with 64GB of Ram compared to 32GB Ram? Currently running 32GB, but assume that's only really effective if caching a slower harddrive? It's a shame that in order to use more than 32GB you need a bigger version of windows rather than just Standard edition.... Out of interest why can we not use StarWind with win 7?

With regards to CPU, would a faster CPU be better than one with more cores? I can never seem to max out the CPU despite the network and SSD not being maxed out. even with caching turned off benchmarking is about 70% performance of native. any thoughts?

Kind regards..

Thu Dec 27, 2012 9:31 pm

We don't support non-server OSses for production as majority of people don't use desktop OSes on servers as we do test most common patterns in the lab.
You can run evaluation on Windows 7 / 8 however.

sanity wrote:Ok, so turns out you can install the service on Windows 7.

What would be the advantage of installing on windows server 2008 r2 instead of windows 7? Win 7 Pro supports up to 192GB ram and costs about £100 whilst 2008r2 costs more than £800 to support more than 32GB ram...

thanks.