ram memory 4gb HDD Performance x2 Not That Simple: Dual Head Arm Hard Drive Test Analysisadmin
ram memory 4gb HDD Performance x2 Not That Simple: Dual Head Arm Hard Drive Test Analysis
disco SSDSSD OEMram memory 4gb
ram memory 4gb+8618750919058, I saw this year’s OCP Global Summit pdf data updated part of the ram memory 4gb, this time I share a total of 47 documents in the second part:
2019 OCP Global Summit – Part2 Speech Materials Network Disk Package Download
extraction code: fkb2
note: You can also click “Read the original article” at the bottom of this article to go to the download page.
combined with the last shared “ 2019 OCP Global Summit Conference Materials Download – Part 1” After extracting, the two parts of the directory can be merged directly.
let’s talk about a more “long” topic – I just worked with hard drives 20 years ago. Four years ago, during OCP 2015, I also compiled 2 Chinese ppt data:
also recommend 2 other articles I wrote:
Today’s article is a Microsoft technology share on OCP, “HDD IOPS – The Never Ending Story,” specifically about the upcoming release of the dual-head arm HDD Mechanical Hard Drive (Multi Actuator Drivers).
dual head arm acceleration principle: single disc 2 LUNs work independently
as shown above, Seagate and WD have prepared this product, and their hardware structure is similar. The existing hard disk is 2 read and write heads on the front and back of each platter, if the 8-disc HDD is a total of 16 heads. Taking this as an example, instead of a dual magnetic head arm design, the heads of the top 4 platters will be driven by the head arm 1, while the magnetic head arm 2 is responsible for the 8 heads of the 4 platters below. See the figure below:
If the upper and lower parts work independently in parallel, it is not difficult to understand that the theoretical performance of the dual magnetic head arm HDD is close to doubling. But there is no free lunch, and adding a magnetic head arm requires an extra motor, bringing about an increase in power consumption and cost; At the same time, there are challenges to the design of the (chassis) mechanical structure, which I will say later.
since the After SSDs, high-speed (10k, 15k) hard drives are getting worse and worse because their previous high IOPS and low latency advantages have become almost worthless in front of flash memory. The mainstream of the future HDD market is 7200 rpm or even 5×00 rpm, but as the capacity of a single disk increases, the average IOPS performance per terabyte is getting lower and lower, so it can only be used for NL nearline storage, such as: photos/videos, data analysis (such as Hadoop), machine learning, etc., and cold data archiving.
there is a curve of “Dual Actuator” in the above figure, somehow the IOPS improvement does not seem to be doubled on the punctuation of 14TB? Let’s still refer to the specific test data in the following article.
I think this schematic is fairly clear: each head arm is responsible for half of the platters has its own The channel and servo systems, although they both communicate with the host through the cache of the hard disk and a single SAS lane, map out 2 LUNs each. If I’m not mistaken, the system will see “2 separate hard disks”, which may seem a bit rough, but it is the simplest and most effective way to improve performance.
thinking question: Suppose you use a dual magnetic head arm hard disk on an ordinary server/workstation, how should you recommend configuring it? Of course it is very likely that you will need to connect to a SAS RAID or HBA card.
early performance test results: slightly lower random writes, sequential write QD1 down?
double-click to enlarge the picture, the same below. The abscissa coordinates are 4KB (1-64 queue depth), 8KB… 4MB, 8MB, I remember that the SATA specification queue depth is 32, and it is necessary to use the SAS interface here.
earlier test results quoted above from Seagate MACH.2, it can be seen that the random read performance of the two LUNs of the dual magnetic head arm hard disk can be maximized at the same time, so that the total IOPS is doubled from the maximum 200 to a maximum of about 400.
Random writing of dual-head arm HDDs seems to be a bit slower, with each LUN up to IOPS at more than 150. In contrast, the write IOPS of a traditional single-head arm hard drive will also be a little slower than reading.
Sequential read performance is also quite ideal, LUN0 and LUN1 each have more than 250MB/s, which adds up to 500MB/s, theoretically 6Gb/s SATA is enough, but with SAS (12Gb/s) reason I just said.
The maximum sequential write bandwidth of a dual-head arm HDD is similar to that of a sequential read, but the performance of low queue depths (mainly QD=1) appears to be poor. I wonder if this will be the case after the official listing?
system design points, how to use the effect well?
let’s look at the system design points 。
– First, due to the increased power consumption of the dual-head arm hard disk, the heat dissipation and power supply requirements of the chassis have increased;
– Secondly, each disk can provide higher sequential streaming (bandwidth), does the network interface speed of the storage platform match? Is there enough upstream bandwidth for SAS Expander?
– J2010 disk enclosure in Microsoft’s Project Olympus project, has been validated to support these new high-performance hard drives.
the AV in the image above is the abbreviation for Anti-Vibration, right?
Olympus J2010 is a 4U 88-disc (3.5-inch) JBOD chassis that can power up to 13W per hard drive. Microsoft said it will face dual magnetic head arm hard drive tests to confirm RV/AV performance. The RV here should refer to rotational vibration, and in addition to the introduction in the recommended link at the beginning of this article, I also compared the vibration resistance of hard disks in real environments ” has been specifically discussed on this topic.
Exptiary reading: “ how Facebook reduced hard drive performance losses from 90% to 2%.
Finally, the speaker pointed out that multi-head arm hard disks alone cannot be used as a solution, that is, they must cooperate with the design of storage systems, which is basically limited to enterprise/data center applications. Microsoft also said it still needs to reduce long-tail latency and needs to work with hard disk manufacturers on IO prioritization.
back to the thinking question I posed above, the simplest way to use it in my head:
– for example, I have a multi-head arm hard disk, 2 LUNs directly do a RAID 0?
– for example, I have multiple multi-head arm hard disks, make one set of RAIDs for each HDD LUN 0, and the rest of LUN 1 for another set of RAIDs, or cross 2 LUNs for the entire RAID10/50/60 or whatever? The problem to consider is that if only 1 LUN of a disk fails, when the replacement hard disk Rebuild is rebuilt, the 2 LUNs can only be moved together.
– multiple copies, EC erasures, and so on….
I finally hope that what I wrote is helpful to everyone. There are still some things to be determined, always follow the development of the industry to learn, this is where the technology is interesting 🙂
note : This article only represents the personal views of the author, and has nothing to do with any organization, if there are errors and deficiencies, please feel free to criticize and correct in the message. To further communicate the technology, you can add my QQ/WeChat: 490834312. If you want to share your own technical dry goods on this public account, you are also welcome to contact me 🙂
respect knowledge, please keep the full text when reprinting, and include the bank and the following QR code. Thank you for your reading and support! Enterprise Storage Technology WeChat public account: HL_Storage
historical article summary: disco SSDSSD OEMram memory 4gb