
Shane Downing Social Links Navigation Contributing Writer Shane Downing is a Contributing Writer for Tom’s Hardware, covering consumer storage, PC hardware, and AI.
bit_user Thanks for the review! I haven't yet finished reading it, but wanted to comment on one particular aspect: So, this is basically what I saw with my Optane P5800X, which is that you can't come even close to saturating the interface with 128 kiB transactions at QD=1. QD=8 will do it. I don't recall how it did with QD=4, but I think that wasn't quite enough. What's weird about it is 128 kiB is the largest transaction size the drive supports! I think my Samsung 990 Pro supports either 1 MiB or 2 MiB transactions, in which case you can basically saturate the interface with QD=1. Reply
thestryker I think it'd be interesting to see something like the Pascari X200Z tested (or dust off a P5800X) just to see where something latency optimized stands. It seems like we're getting a lot more variance on enterprise SSDs these days with regards to segmenting by performance. Reply
Nikolay Mihaylov bit_user said: So, this is basically what I saw with my Optane P5800X, which is that you can't come even close to saturating the interface with 128 kiB transactions at QD=1. QD=8 will do it. I don't recall how it did with QD=4, but I think that wasn't quite enough. What's weird about it is 128 kiB is the largest transaction size the drive supports! I think my Samsung 990 Pro supports either 1 MiB or 2 MiB transactions, in which case you can basically saturate the interface with QD=1. Given 128kiB max request size, if you pass a 1MB of data to the OS to write to an NVMe, it should queue 8 requests for it. And Q8, you say, is ok. Q1 is for scenarios where you can not know the content of the next request (e.g. what to write or where to read from) untill the previous one is over. Reply
bit_user Nikolay Mihaylov said: Given 128kiB max request size, if you pass a 1MB of data to the OS to write to an NVMe, it should queue 8 requests for it. And Q8, you say, is ok. Well, I tested using unbuffered I/O via Linux' Flexible I/O tester which does low-level transactions and don't get segmented. So, any request size larger than 128 kiB failed. Then, while querying the device, I think I saw some NVMe attribute which actually stated the maximum request size was 128 kiB. At an OS level, I also wouldn't trust it to overlap the segments of an application-level request. Maybe there's an I/O scheduler parameter you can set to make it do that. Nikolay Mihaylov said: Q1 is for scenarios where you can not know the content of the next request (e.g. what to write or where to read from) untill the previous one is over. I know what you're saying , but QD=1 sequential is a common test case. Reply
Key considerations
- Investor positioning can change fast
- Volatility remains possible near catalysts
- Macro rates and liquidity can dominate flows
Reference reading
- https://www.tomshardware.com/pc-components/ssds/SPONSORED_LINK_URL
- https://www.tomshardware.com/pc-components/ssds/phison-pascari-x201p-15-36tb-ssd-review#main
- https://futureplc.com/terms-conditions/
- Nvidia amassed $530 billion in financial guarantees to corner the AI market
- Grab a beginner-friendly Anycubic 3D printer for under $300, back at its record-low price
- Lambda Boosts AI Factory Resiliency and Efficiency With NVIDIA Photonics
- Grab a beginner-friendly Anycubic 3D printer for under $300, back at its record-low price
- Grab a beginner-friendly Anycubic 3D printer for under $300, back at its record-low price
Informational only. No financial advice. Do your own research.