如何选择硬盘尺寸？(What size drive should I use?)

如何选择硬盘尺寸？

Neil Cameron

        一言以蔽之…关键在于选择正确尺寸的硬盘来适应不同的性能需求。目前，许多人只是基于容量来选择硬盘的数目以及所需的总容量大小，但是，涉及到硬盘容量和性能的选择，还有若干因素需要考虑在内。我将分两大类来进行讨论——旋转介质和SSD：

旋转介质：

      总的来说（这句话势必会给我带来麻烦），旋转介质的性能基本上是差不多的，不论其容量大小。比如Seagate的Constellation 2 SATA硬盘，该公司给出的数据是寻道、读、写的速度完全一样：对容量高达2TB的硬盘而言，速度都是250GB。所以，容量的增加并没有带来相应的速度提升。

        这一点为什么很重要呢？在由旋转介质组成的RAID阵列中，转轴的数目对阵列的性能存在重大的影响。例如，一对镜像盘（两块6TB的硬盘 —— 6TB容量）的速度就远远比不上RAID 5阵列中8块1TB硬盘（大约7TB容量）传送数据流的速度。RAID 5阵列中额外的转轴为每块硬盘提供了小规模的读和写，从而大大提升了速度。

       来对传送数据流进行些粗略的计算——如果每块硬盘的读速度是120MB/秒，那么，理论上来说，从镜像盘得到的最高速度为240MB/秒，而从RAID 5得到的最佳结果则为840MB/秒。当然这都是理想化的数据，但其间差距可见一斑。

       当然，RAID配置也很重要，需要与客户将要应用的数据类型相匹配，原因在于，总体来讲，对于旋转式介质而言，转轴愈多，性能就愈高。虽然还有能耗和成本以及其他方面的考虑，但我这里主要考虑的是性能，那就让我们专注于此。

SSD :

       对SSD而言，小SSD盘与大SSD盘的性能指数确实存在着巨大的差距。其间当然还有成本的差异，来看看具体数据吧…

       先来看看Intel的DC3500SSD（非常、非常优秀的一款产品），80GB和800GB的硬盘的IOPs的差别并不大（分别是70K到75K不等），但顺序读写的速度就有非常显著的差异。80GB硬盘的顺序读速度据称达到了340MB/秒，而800GB硬盘的顺序读速度则达到了500MB/秒。顺序写的差距就更加惊人，80GB硬盘的速度达到了100MB/秒，而800GB硬盘的写速度则宣称是450MB/秒。

       这些到底意味着什么呢？

       在任何RAID阵列中，转轴的多少（在此即指更多硬盘）会带来性能上的不同。如果控制卡能将读/写任务分配到多个硬盘，那么控制卡向每块硬盘读/写的数据就减少了，因而可以更快地完成任务。再加上大硬盘的速度要快上许多，那么改善效果就加倍突出。

       难点在于，硬盘数目不多的阵列中，会遭遇RAID的局限，即硬盘读写的数据大小比分摊到更多盘片时要大得多，因此需要找到一个平衡。

        RAID 0中8块80GB的硬盘最高可以提供640GB的容量，读速度可达2720MB/秒；5块240G的硬盘可以提供大概同样的速度，容量则为1.2TB。

        故此，直接提高硬盘的大小，期望能实现相同的容量，并维持性能水平是不切实际的。RAID盘阵的基本思想决定了盘的数目依然对速度起到了决定性影响。

       真正的问题还在于性能测试，也正是这一点引起了我的注意。许多人对硬盘速度具体意味着什么缺乏概念——只是看到宣传中的夸大信息，晓得Intel的盘“最高”能提供500MB/秒的读速度（这也是我从宣传材料中汲取的信息）。

        因此，当测试八块 DC3500 SSD时，我本应该能在RAID 0中得到4000MB/秒的读速度，对不对？答案应该是肯定的…… “是啊，如果你用的是240GB或者更大的硬盘时确实如此”。但是，如果硬盘只有80GB或120GB的容量，那么最高能达到的速度就分别是2720MB/秒和3560MB/秒（理想情况下）。

       这样的结果会让你向RAID厂商咆哮一番，责备他们提供的设备性能不够好……没能充分发挥硬盘的性能！而事实上，是硬盘在拖控制器的后腿。如果在控制卡上接上足够多的速度足够的SSD，Adaptec的7系列和8系列控制器可以传送数据流，跑满PCIe总线（约6600MB/秒）。但是，就算在控制器上连上16块80GB的Intel DC3500硬盘，还是会发现性能无法达到极限，原因很简单，这些硬盘的速度并不如想象中的快。

结论

       这些与我们有什么关系呢？眼下真有人在用全闪存的系统吗？回答是肯定的，而且用户数目在不断攀升。SSD在各种规模的企业及消费者中的应用已经日渐成熟——现在对全闪存系统的认识也趋向准确，大家认为此类设备的可靠性、速度及性价比都很不错，并且能为整个计算机系统带来焕然一新的改变，因此至少是值得慎重考虑的一种选择。

       只是，需要确保这些系统的配置正确合理，并保证购买的系统的速度达到要求——事实上，所有厂商的SSD都有同样的配置的问题。

 

What size drive should I use?

In one word, this is all about the issue ofchoosing the right size drives to suit your performance requirements. Now a lotof people just choose the number of drives based on capacity, and what overallsize they need, but there are a few considerations that should be looked atwhen it comes to size of drive and performance. I’ll break this into twosections – spinning media and SSD:

Spinning Media …

As a general statement (bound to cause me somegrief), the performance of spinning media is pretty much the same across afamily of drives, no matter what the size. If you look at something likeSeagate’s Constellation 2 SATA drives, they quote seek, read and write speedsof exactly the same numbers for 250GB up to 2TB drives. So while they getbigger, they don’t get faster.

So why does this matter. Well, in spinning mediaRAID arrays, generally the number of spindles has a major impact on theperformance of the array. For example, a mirror (2 drives of 6TB drives – 6GTBcapacity) will not be anywhere near as fast in streaming data as 8 x 1TB drivesin a RAID 5 (7TB capacity or thereabouts). The additional spindles in the RAID5 allow for small reads and writes from each drive, speeding things upconsiderably.

To do some really rough maths on streaming data –if a drive does 120MB/sec read speed, then theoretically the best speed you canget from the mirror is 240MB/sec and from the RAID5 is would be 840MB/sec.That’s all in a perfect world of course, but you get the idea.

Of course, the RAID configuration matters and needsto suit the data type you are building for, but in general, with spinningmedia, you can say that more spindles equals more performance. Yes, there arepower usage considerations, and cost considerations (though not a great deal),and those all need to be taken into account, but I’m talking about performancehere, so stay focused on that side of the equation.

SSD

On the SSD side of the equation, there is in fact abig difference between the performance numbers of a small drive vs those of alarge drive. Yes there are cost differences as well, but let’s look at thenumbers …

Looking at Intel’s DC3500 SSD (a very, very goodproduct imho), there is not a great deal of difference in the IOPs speed fromthe 80GB to the 800GB drives (70K to 75K respectively), but in the streamingspeed there are some pretty dramatic differences. The 80GB drive claims asequential read speed of 340MB/sec, while the 800GB drives claims a sequentialread speed of 500MB/sec. The write speed difference is even more dramatic, withthe 80GB drive writing at a claimed 100MB/sec, and the 800GB drive writing at aclaimed 450MB/sec.

So why does this matter?

As in any RAID array, more spindles (or in thiscase drives) matters. If the controller card can split the reads or writesacross multiple drives then it reads or writes less data to each drive,finishing tasks quicker. If you add to this the fact that the larger drives aredramatically quicker, then the effect is multiplied.

The conundrum here is that a small number of driveswill still suffer from the RAID limitation that the reads and writes to thedrives will be larger than it would be if they are spread out across a lot moredrives, so a balance is required.

8 x 80GB drives in RAID 0 would give a potentialread speed of 2720MB/sec at 640GB capacity.
5 x 240GB drives would give pretty much the same speed, at 1.2TB capacity

So you can’t do a straight upgrade of drive size tocome up with the same capacity numbers and still maintain performance – themathematics of RAID still means more drives equals more speed, but you start tosee the point.

The real problem here lies in performance testing,and this is what brought this to my attention. A lot of people don’t read thefine print regarding the speed of the drives – they just look at the marketingblurb and see that the Intel drives are capable of “up to” 500MB/sec read speed(that’s pretty much the way I read the marketing material as well).

So when testing 8 x DC3500 SSDs, I should be ableto get 4000MB/sec read speed in RAID0, correct? The answer to that is “yes … ifyou are using 240GB or larger drives”. However if you only have 80GB or 120GBdrives, then your maximum speeds will be 2720MB/sec and 3560MB/sec respectively(in a perfect world).

This will have you yelling at your RAID vendor thattheir performance is not good enough … it’s holding back the drives! Whereas inreality it’s the drives holding back the controller. Adaptec’s 7 and 8 Seriescontrollers can stream data to the full extent of the PCIe bus (approx6600MB/sec) … if you put enough of the right speed SSDs on the card. Howevereven if you put 16 x 80GB Intel DC3500 drives on your controller you’ll be shyof that performance benchmark simply because the drives are not as fast as youmight have thought.

Conclusion

All of this matters because? Are people reallyusing pure SSD systems now? Well yes, they are, in ever-increasing numbers. TheSSD has come of age in the mind of the enterprise, small business and consumer– they are now quite rightly regarded as reliable, fast and reasonably-affordabledevices that make such a dramatic difference to a computer system that they areat least worth considering.

Just make sure that you are configuring themcorrectly, and that you are in fact sure of what speed you are actually buying.This has not been a bash against Intel – all SSD from all vendors suffer thesame configuration issues.