|
4 Benches with a RAID 0 of CF cards connected on Sata.
We then went into the same bench tests with this time a Sata RAID 0 of Extreme IV 8Go cards. In order to acheive this, IDE to Sata adapters will be connected onto our CF to IDE adaptors, actually Serillels 2 from Abit. Adding up adaptors can deteriorate performances, but it is not the case here : on a one CF system, only a small CPU usage increase (see belows) is noticeble between Sata and IDE.
We chose a RAID 0 since on Compact Flash we don't have the same risks of data loss as on a hard drive, due to the absence of mechanical parts. RAID 0 is a reliable solution and presents very little risks compared to a RAID 0 of hard drives...
In the configuration of our RAID, the bios of the RAID card (integrated to the AN7) allows us the use of different strip sizes : 4, 8, 16, 32, 64, 128Ko. The choice of the strip size of the RAID is essential, as it greatly influences the performance. Depending on the type of file hosted on the server, it will enhance or lessen performances...The choice can only be done during the RAID creation, and cannot be changed without breaking the RAID : So, be careful!
Functional principle of RAID 0 and strips : each file written on the RAID will be cut in many pieces with the strip size we defined in the RAID creation.
Examples:
A 750 Ko file must be written on the disk.
1 16 Ko strip size: The system will write 750/16 -> 49 strips, which gives us 25 accesses to the CF to write 16 Ko each time.
2 128 Ko strip size: The system will write 750/128 -> 6 strips, which gives us 3 accesses to the CF to write 128 Ko each time.
Let's now take a 15 Ko file
1 16 Ko strip size: The system will write 15/16 -> 1 strips, which gives us 1 accesses to the CF to write 16 Ko each time.
2 128 Ko strip size: The system will write 15/128 -> 1strips, which gives us 1 accesses to the CF to write 128 Ko each time.
That's for the RAID principle, but things get more complicated due to the writing management of the CF, as it is done by page. In order to write a 2 Ko page, the cluster where the page lies will have to be entirely rewritten, needing 128 Ko to rewrite. When we add that to the RAID strip control, it becomes clear that the strip size choice isn't an easy task...
Now that we've seen the principle, we can notice that large strip sizes may penalize performances on small files. The solution doesn't necessarily lie in the use of small strip size as with most bottom-of-the-line RAID controllers. It's the machine's CPU which is in charge of cutting up the file into strips. So, the smaller the strip size, the busier the CPU management.
In the best of worlds, the perfect strip size would be the one cutting in half a maximum number of files (for a two unit RAID). In that case, a controller makes an access followed by the writing, on each disk, giving near to double performances. If, for the same file, we cut it in four blocks, it gives us four accesses and four writing operations. The rate being the same, writing 16 Ko once or 8 Ko twice will take the same time BUT the access time will be doubled! It's in those cases that the CF, with its minute access times, sees its performances enhanced : the access times being 30 times faster, multiplying accesses won't be as penalising as on hard drive.
That was for the theoretical part. Let's now see how it goes with practical tests.
We have, for each test, taken values with each possible strip size : 4Ko, 8Ko, 16Ko, 32Ko, 64Ko, 128Ko.
Clearly during the tests, strip sizes of 64 Ko and 128 Ko make the performances collapse in a way that RAID amateurs know very well! Using 64 Ko, performances fluctuate between our maximum in RAID and a non-RAID system (here in blue). The Burst Speed stays about the same, but, as predicted, the CPU usage is higher in RAID's with smaller strip sizes.
We can observe the same as previously, but can see the CPU usage with more precision, rising as the strip size drop,as said in the introduction. On the performance side, the "deep drop" starts a bit earlier and 32 Ko strip size already displays a drop compared to 4, 8 and 16 Ko strips.
With only one Compact flash we had a system wich seemed already more reactive, going into RAID conforts us in this idea. It is to say the system put in place (RAID 0 of Extreme IV) is faster than all the SSD disks actualy around.
Our RAID 0 Sata of Extreme IV CF cards allowed us to obtain an SSD disk of 16 Giga with a transfer rate in reading of 65 Mo/s minimum,and an access time of 0.5 ms.For 300€,our RAID is faster and cheaper than commercial SSD disks.
|
- Vous pouvez renouveler le code de sécurité en appliquant un rafraîchissement à votre navigateur.
|
|
)
)
)
)
)
)
)
)
)
)
)
)
Lire la suite et commenter cet article
