[sp332]

That's within an order of magnitude of RAM. You're getting 16TB of RAM!

[altcognito]

> This volume can host a 16 TB and 20,000 IOPS I/O-intensive transactional database requiring single digit millisecond latencies with consistent performance using Provisioned IOPS (SSD) on AWS

No, not even close.

[vidarh]

I think you severely underestimate RAM bandwidth on modern servers.

[sp332]

OK, maybe not top-of-the-line RAM, but some servers do have RAM that slow!

[ak217]

PCIe or SATA attached SSD random access latency is around 0.2ms. Typical RAM latency is 100ns, maybe 200 in a NUMA cross-node access. That's a 3 orders of magnitude difference. Add another order of magnitude for network-attached SSD.

Bandwidth wise, a single DDR3 channel has around 10 GB/s (and a typical server has 4 to 8 of them). A single half duplex 10GE link (the most you can provision and effectively use on EC2) is 500 MB/s. So, generally 1 to 2 orders of magnitude.

[sp332]

Ok the latency is high. But I was seeing numbers in the 300,000 IOPS range for ramdisks.

[ak217]

A naive guess is that most of that is filesystem overhead. Were you using tmpfs?

[sp332]

It wasn't my benchmark, but that's my guess too.

[mrmondo]

Actually it's considerably slower than the single SSD in the lowest model MacBook Pro.

[sp332]

Really? Over 20,000 IOPS?

[mrmondo]

Yes - they do around 850MB/s write and 80-120K IOP/s.

Hell, even my 3 year old OCZ vertex 4 in my old 2011 MacBook Pro does 500MB/s write and 65-80K IOP/s