前回, Phoronixで紹介されたが, 今回はOpenNet.ruとかいう良く分からんサイトで紹介された.
ロシア語はわけが分からないのでやばい. そこでGoogle翻訳にかけてみると完璧な英語になった. すごすぎる. 内容を見るに, LKMLやPhoronixの情報を自分の言葉でまとめ直しているだけだろうと思う.
Linux kernel is a realization of a new caching mechanism for SSD-drives - dm-writeboost. The main difference between the new system from the existing arrangements DM-Cache and Bcache is focus on packet writing to extend the life of SSD-drive and speed up writes. The module dm-writeboost claims for inclusion in the experimental staging tree next release kernel Linux. Dm-writeboost stores the information in a log-like series filled cyclical structure to fill that random set of write operations is collected and discharged to the SDD in batch mode as a large contiguous block. This approach will significantly increase the efficiency of recording - using dm-writeboost loss of performance caching random queries is only 3% compared to the sequential one. For example, if the capacity of SSD-drive to 266MB / s, dm-writeboost provides write speeds of 259MB / s.
コメントもついてる. 番号がついてるのでその順に翻訳. 意味分からん. ちゃんと翻訳出来るのは, まともなロシア語に限るらしい. どちらかというと批判されてるような気がする.
- Er ... Asynchronous write cache memory? What for it is, is no train reliability.
- On statsionarnike without UPSA - yes. Is targeted, as it is for every compact Soup with vsotoennoy UPS system.
- but how many of these drives that have their hardware architecture is not open? And then how much worse and more difficult and the development of open source software will be obtained for them? (to the extent such chudotehniki abroad are cheaper than here in Russia?)
- A sense of a separate module? Magazine in the frame-ROM or soar brain. :)
- Hmm, the author forgot to mention how much memory it occupied the cache ...
- Why in the text news from SSD-drive 266mbayt / s instead of 6Gb / s who write in advertisements and product characteristics?
- Guess who better to trust on the speed of the drives: marketers or driver developers? Hint - the salary is much weaker than the second depends on the sales.
- 6Gb / s - it's bandwidth SATA / SAS-bus. 266mbayt / s - this is the writing speed to the memory cell.
- Before we get to the block device, the data is cached in RAM (unless of course you are not forced to discard them using fsync ()). Then add another layer of caching on the SSD. Of course, the reliability will decrease: for all cached write operations on the block device, it stops working and logging in case of power supply failure to the lost data, but can be broken structure of the file system. However, if you are worried about the safety of the data is, you should in any case ensure uninterrupted power supply. Or do all the time fsync-and multiple corresponding loss of performance.
- if the record is consistent, then why SSD, 10Krpm drive will be enough.
- Well-BE SSD can drop stronger. Eating less, weighs less. Easier to do! In HDD dofiga metalworking and fucking complicated system servomehaniki. Do not see what made simple. Super accurate in the first place, it is difficult to control in the second.
- The structure of the FS can not be compromised for the write operation is atomic. Just part of the data will not be recorded, along with the magazine, and of course will sit on the ssd. Actually reduce reliability should not be
- like head movement and the time it has not been canceled
- HDD is cheaper, and it outweighs all of that stuff you've written
- Marketers such marketers.
- Ingeniously, Watson. And it is better to data on the ramdisk. It is true then, after fakapa have to soar brain matter "Oh, and where it all go?" :)
- Then that SSD generally much faster. How much do 10k RPM generally squeezes under 300 megs per second? There is also at least RPM and above, but no recording density, so that the transmission rate on pancakes is not so fabulous.
- Plus the mechanics potentially unreliable and may decline catastrophically, for example, surface scratch - and all the data goodbye. On the other hand flash degrades gradually and overall drive - relatively predictable.
- The structure of the FS is not broken not because of atomicity, and because of barriers record.
- Only now suddenly die anyway :)
- It is there and there is not faster niterfeysa. 10K RPM SAS vs SATA-ii ... anything can be.
- Yes, they are dying because of the controller.
- The battery flew / glyuknul all.
- Just for the options with which data can give a shit, usually written PROBABLY LOSS OF DATA. And ask again 3 times. And so it is clear that the servers are 2 power supplies that are stuck in different UPS plus battery on the raid controller. For example, when the writeback on a raid controller, even with a battery, GUI controller still will clarify the "are you sure?".
- Sdohshih SSD (the devil's work to put the cache and database cache SQUID) for the last 2 years, I have 8 pieces. But sdohshih HDD - none. Question: what happens reliable, HDD or SSD? I have a HDD on the servers go for 5-7 years (it's hard now to get SCSI drives 68-pin, on order only), SSD, more than a year not one handed.
- Read then the cache is not in a batch mode. And in a very chaotic.
- 200 issue, no problem, put 2 to 400.
- Well hello, linear writing speed was measured 15krpm?
- Have you tried to scrap a sawmill?? The cache SQUID - buy 4-ECC DDR3 bars. But Reid on ssdshkah a base sikvelservera gives screws on Balde from 6 to 11 times!
- mete, SSD - faster. pipiski: CDM 15.7k Seagate 300G in the mirror and Intel SSD 520 (like 240 G) as in a mirror. Guess we will not, intel won.
Although the gap was not as enormous as iopsah on the same media on the database. As I wrote above - on the processing of requests received severe tests from 6 to 11kratnogo increase execution speed.
dm-writeboostは万能ではない. ユーザランドの世界では, これ使っておけば80%のケースでは問題ないよということが多いだろう. しかしディスクキャッシュの適用は, ともすればただの足手まといになる. 自分たちのストレージを適切に分析して, 適切なキャッシュソリューションを選ぶことが重要であるが, それはわりとスキルのいることだろうと思う. とにかく試してみて, 結果をブログでも良いので公開して欲しい. 分からないことがあればTwitterでもメールでもすべてウェルカムだ.
こいつら何もわかっちゃいねえ. まず, PDFくらいは読め. 想像だけで話すのは良くないな.