Making The Case For Flash Over Disk In Power Systems
May 2, 2016 Timothy Prickett Morgan
Two weeks ago, concurrent with the launch of IBM i 7.3, Big Blue announced a slew of new flash storage drives and storage controllers, which we told you about based on the announcement letters as we normally do. Since that time, we have got our hands on some internal analysis that IBM has done for Power Systems partners and customers, and are sharing that with you to help in your buying decisions as you contemplate adding flash to your systems or, perhaps, even going all flash. Enterprises are doing it, now that the price for effective capacity after data compression and deduplication for flash within many arrays has in many cases reached parity with disk-based controllers. To the best of my knowledge, there is no such compression or de-duplication on native storage within IBM i servers, but a fair comparison would also take into account that many IBM i shops don’t buy their disk drives for capacity, but rather for a certain number of disk arms and therefore a certain I/O operations per second rate and in many cases disk drives are short-stroked. This means the drive is not allowed to get full and only stores data on the outside edge of the disk to get the maximum rotational speed of the platter and therefore a higher IOPS for the drive. IBM’s FlashSystem external arrays have these features, by the way. But in many cases, data compression and de-duplication may not be necessary to justify a move to flash for IBM i shops, who have the benefit of single-level storage thanks to the operating system if they use directly attached storage inside their system enclosures. This is particularly true as IBM expands the capacity of its flash-based SSDs with its fourth generation of enterprise MLC drives, which come in 2.5-inch and 1.8-inch form factors. Here is how the 2.5-inch SSDs have evolved over the years: And for fun, here is a performance table that shows the evolution of the performance of the eMLC SSDs over time and compares them to 15K RPM SAS disk drives in the same 2.5-inch form factor and the new read-intensive SSD: The performance specs above were based on a Power Systems machine running with a single drive and at a sustained rather than burst I/O load. So take them with a grain of salt. The thing to note is that the capacity, IOPS, and throughput of the eMLC flash SSDs just continues to get better and better and as we all know, disk drives are stuck in the mechanical mud. You can make a disk drive more capacious–8 TB drives are now common, and 10 TB and 12 TB drives are on the way–and you can make them cheap, but you cannot change the laws of physics. In the table above where performance figures have an asterisk (*), the lower number was tested using a PCI-Express 2.0 SAS adapter, while the higher number was using a PCI-Express 3.0 SAS adapter with more oomph. The percentage improvements in capacity and performance are for IBM’s eMLC3 drives versus the new eMLC4 drives. The eMLC4 drives are available as standard units that are designed for a mix of read and write operations, and there is also a read-intensive, or RI, variant of the 2.5-inch SSDs that have a limited amount of writes they can do before they start degrading and losing their effective capacity. In general, with the 2.5-inch drives, the 387 GB and 775 GB drives have been refreshed with better NAND flash, which is less expensive for IBM to buy on the market and now less expensive for customers to buy and put into their systems. For instance, using the eMLC3 flash, a 387 GB drive cost $3,588, but the new eMLC4 drive in the same capacity now costs $2,399, which is a 33 percent price cut. If customer want to move to the new 1.55 TB SSD, that offers a lower cost per GB, and the new 1.8 TB read intensive SSD (which only comes in the 2.5-inch form factor) has the lowest cost per GB that IBM has ever had for flash storage on Power Systems, according to the IBM guide. If you bought the 775 GB eMLC3 drive and used it on read intensive jobs, it would cost you about $8 per GB, and it was not designed for heavy reads. The new 1.9 TB read intensive drive costs around $2.4 per GB, which is 70 percent lower per unit of capacity. (This assumes 4 KB rather than 528 B sector sizes. You pay a slight premium for 528 B sector sizes.) Maintenance on the new eMLC4 drives is anywhere from 40 percent to 50 percent lower, too, which is important over the long haul. Take a look: So here are some real world comparisons that IBM cooked up to think about these new 2.5-inch SSDs and how they compare to disks and prior flash drives. Say, for instance, you have a Power S824 machine that you are going to keep for five years (two years beyond the warranty). With 14 disk drives with 146 GB capacity (139 GB running IBM i) in a RAID 5 configuration for data protection, you would get around 1.8 TB of usable capacity, but due to short stroking and to get an appropriate amount of I/O for performance, you would end up with about 900 GB of usable capacity for around $6,972. With the third generation flash SSDs, four drives at 387 GB with RAID 5 protection would give you about 1.15 TB of usable capacity, and you could use all of it, for around $16,080. Unless you need high IOPS, this was probably going to be too expensive for your blood. Now, enter the new eMLC4 SSDs. Four of them give the same 1.15 TB of usable capacity with RAID 5 data protection, but now it costs $10,100. It is only 45 percent more expensive and has much higher IOPS than the disks. And if you do the comparison with disks, which cost $7.75 per GB (assuming short stroking), the SSDs at $8.78 per GB are really only 13 percent more expensive on a unit of capacity basis. And you get a heck of a lot more I/O. The flash drives will last five years, like a disk, so don’t sweat that unless you have heavy write environments. Let’s take a high-end example to bookend this. You have a Power E870 server and you are looking at adding 144 drives of the same capacity to six EXP24S drawers for around 10 TB of usable capacity with RAID 1 mirroring turned on for data protection. This will cost you around $264,708 at list price. With the eMLC3 SSDs, 36 drives at 387 GB using RAID 5 data protection would yield around 14.7 TB of usable space in two EXP24S drawers, but it would cost $325,958–about 23 percent more. This is a harder sell, even with the higher IOPS, lower heat and noise and electricity consumption. Now, move to the eMLC4 drives, and with the same 36 SSDs, the price drops to $203,774 and that is 23 per cent less expensive than the disk drives for 47 percent more capacity. You could drop that flash price a lot lower by getting the same capacity as the disks, of course, but IBM did not think of that. But it should be around $145,000, which is 45 percent lower. Yes, flash is less costly than disk in this example, and offers a lot more throughput and IOPS. In this scenario, you could also almost double the capacity of the flash array based on eMLC4 drives to around 19 TB and spend only $250,550 on the arrays–still less than the disks cost. These comparisons assume there is no short stroking for performance, but if there is, the gap between flash and disk will double–and not in the favor of disk, even on a capacity basis. The question now becomes why anyone buying a Power8 machine to do transaction processing will use disks at all. In addition to the 2.5-inch SSDs cited above, IBM is also shipping updated 1.8-inch drives, which are used in two-socket and four-socket Power Systems machines based on the Power8 processors. Here are the feeds and speeds of these units: And here is how to think about the old and new generations of 1.8-inch drives: The 1.8-inch drives cost the same as the 2.5-inch drives, so on a cost per capacity basis they are interchangeable as far as the operating systems are concerned and the same kind of comparisons will apply. The IOPS and throughputs for the 1.8-inch units are a little different from the 2.5-inch SSDs, but not hugely so. The same kind of thinking will apply. One last thing: IBM i has a feature called Fuel Gauge for the read intensive 2.5-inch drives that will warn customers about when they are running up against the write and read limits of the drive. After the drive has done 3,394 TB of writes (that’s nearly 1,800 full drives of the drive), it will hit its maximum write capacity and performance will degrade and capacity will start to diminish. RELATED STORIES Sundry April Power Systems Announcements IBM i 7.3 Arrives April 15; 7.2 TR4 Follows in May IBM Puts Future Power Chip Stakes In The Ground The Value Of Things That Don’t Change–And Those That Must A Possibly Coherent Future Power Hybrid System What Will IBM i Do With A Power10 Processor? OpenPower Could Take IBM i To Hyperscale And Beyond OpenPower Builds Momentum With New Members, Summit Aiming High, And Low, With Power Chips How IBM Stacks Up Entry Power8 Machines Against X86 Iron Plotting Out A Power Systems Resurgence A Real Open Power Server, Finally Google Reveals Its Own Power8 Motherboard; Can It Run IBM i? What The System x Selloff Means To IBM i Shops Power8 Processor Packs A Twelve-Core Punch–And Then Some IBM Forms OpenPower Consortium, Breathes New Life Into Power
|