It really depends on your workload. If your working-set fits into memory and is read-mostly, your I/O is basically irrelevant. Even if you're not in-memory, if you are doing mostly sequential reads (like the OLAP queries they are benchmarking), each drive can do ~200MB/second, so it is not too difficult to become CPU bound if you are doing any significant processing on each row (particularly if your row size is small!)
For OLTP queries, particularly with working-set > RAM or lots of writes, your disk I/O is probably the bottleneck (probably your IOPS, actually, which is why SSD can be so valuable). Pretty sure they're not targeting that use-case though!
in my experience (which isn't the sum-total of all human experience), our OLAP queries have a bottleneck at the fiber interconnect between the storage controller and the OS. cpu still isn't the problem for our OLAP systems.
your point that great strides have been made in storage such that we could be back at cpu as a bottleneck is well taken though.
It is a reasonable assumption given he said "raid 10". By "cheap" I mean all spinning disks. Enterprise SSDs are an order of magnitude more expensive than the best spinning platters. ~$25k for 3TB.
Raid 10 basically just gives performance over Raid 1. You might Raid 1 SSDs but Raid 10 is likely to be a very expensive way to hit your IO bus limits.
Given that, assuming it is spinning disk is a reasonable assumption, and the accusation of a strawman was unnecessarily rude.
I don't find it to be unnecessarily rude. It was a legitimate critique of the poster's own assumptions. I guess that because I'm not an HN "insider" I'm not allowed to challenge someone's assumptions. Only once you hit the magical karma level are you able to do such things.
Comments
From my experience I/O is more important than CPU for DB, I've never seen the 24 cores of our servers having problems, however the raid 10...
It really depends on your workload. If your working-set fits into memory and is read-mostly, your I/O is basically irrelevant. Even if you're not in-memory, if you are doing mostly sequential reads (like the OLAP queries they are benchmarking), each drive can do ~200MB/second, so it is not too difficult to become CPU bound if you are doing any significant processing on each row (particularly if your row size is small!)
For OLTP queries, particularly with working-set > RAM or lots of writes, your disk I/O is probably the bottleneck (probably your IOPS, actually, which is why SSD can be so valuable). Pretty sure they're not targeting that use-case though!
in my experience (which isn't the sum-total of all human experience), our OLAP queries have a bottleneck at the fiber interconnect between the storage controller and the OS. cpu still isn't the problem for our OLAP systems.
your point that great strides have been made in storage such that we could be back at cpu as a bottleneck is well taken though.
I assume you are using cheap spinning disks, and not something like a fusion-io drive. How big is your dataset?
that's a narrow assumption. why not get more information from the person instead of simply attacking a straw man.
It is a reasonable assumption given he said "raid 10". By "cheap" I mean all spinning disks. Enterprise SSDs are an order of magnitude more expensive than the best spinning platters. ~$25k for 3TB.
not sure why i was downvoted, but oh well.
really? you don't RAID your SSDs?
Raid 10 basically just gives performance over Raid 1. You might Raid 1 SSDs but Raid 10 is likely to be a very expensive way to hit your IO bus limits.
Given that, assuming it is spinning disk is a reasonable assumption, and the accusation of a strawman was unnecessarily rude.
I don't find it to be unnecessarily rude. It was a legitimate critique of the poster's own assumptions. I guess that because I'm not an HN "insider" I'm not allowed to challenge someone's assumptions. Only once you hit the magical karma level are you able to do such things.