 |
|
|---|
Battle of the Titans: Promise SuperTrak 100 vs. 3Ware Escalade 6400
|
It would be an understatement to say that the popularity of RAID has increased over the last few years. It wasn't long ago that those wishing to implement RAID had to go with SCSI drives and controllers. The relatively high cost of SCSI limited RAID, for the most part, to the server arena. Then, in 1997, Promise Technologies introduced the first ever ATA RAID card: the FastTrak (now called the FastTrak-33). The floodgates opened with this introduction... even folks who barely knew what RAID was knew that they wanted it. It had never before been possible to configure a RAID array in such a cost-effective fashion: the FastTrak debuted with a pleasant sub-$200 price tag and the ATA drives, of course, were significantly less expensive than their SCSI counterparts. As the price of ATA drives plummeted, the popularity of ATA RAID grew. Promise followed up on the FastTrak with the FastTrak-66 and FastTrak-100. Other manufacturers - such as Iwill, AMI, and HighPoint - recognized the exploding ATA market and introduced their own cards accordingly. It wasn't long afterwards that motherboards started to integrate ATA RAID chips from these very companies.
Not everything was rosy in the land of ATA RAID, however. The FastTrak series of controllers - as well as offerings from Iwill, AMI, and HighPoint - strove for low cost in an effort to appeal to the desktop market. As a result, it's no surprise that performance issues have been at the forefront of discussion. Some claim that these low-end ATA RAID cards do little for real-world performance, while others insist that significant performance gains are indeed realizable. Furthermore, whether or not these cards are even "true" hardware RAID also continues to be a cause of much debate -- some say that a card must have an on-board RISC processor to be considered a hardware implementation. With all of the above in mind, the introduction of undeniably "true" hardware controllers - complete with on-board processors, several independent ATA channels, a higher price tag, and, in some cases, on-board cache - was inevitable. Indeed, Promise, 3Ware, and Adaptec have since introduced their own line of hardware-based ATA RAID controllers.
Promise's SuperTrak line of cards has recently drawn significant interest from the online community. Although introduced in 1998, the SuperTrak languished in obscurity. A year later came the SuperTrak-66, again without much fanfare. Then, last September, Promise announced the 3rd generation SuperTrak-100. SuperTrak 100 Specifications:
The SuperTrak-100 immediately caught enthusiast community's attention, hardly surprising considering the card's specifications. After all, a 6-channel, ATA-100 RAID 0/1/01/3/4/5 card with 16MB of on-board cache and hot-swap support is quite a feat, especially when the ATA RAID community as a whole had previously been used to much less expensive, less feature-rich controllers. The Card... The first thing one is likely to notice is the SuperTrak100's sheer size. The controller's full-length design permits a board that features six IDE connectors, three ATA-100 ASICs, and an Intel i960 processor. Included with the card are the following items:
The Software... RAID arrays can be created, viewed, and deleted via the SuperTrak-100's BIOS (a.k.a. SuperBuild). Arrays may be created manually (allowing more user control) or via what Promise calls "Auto Setup." The user may also assign one or more drives as a hot spare as well as designating a bootable array. The package also includes software for remote monitoring and configuration of arrays. The utility, SuperCheck, permits creation and deletion of arrays and configuration of several array parameters. Read and write cache may be independently toggled and several cache-policy settings may also be set. These include "flush frequency timer" (the amount of time a block of data written to the cache can remain there until it is written - i.e., flushed - to the drives), "dirty threshold flush start" (when the percentage of dirty blocks in the cache exceeds the threshold, flushing begins automatically), and dirty threshold flush stop (Flushing stops when the percentage of dirty blocks in the cache falls below this threshold.) In this review, all of the above cache policy parameters were left at their default settings (2 seconds, 50%, and 5%, respectively). SuperCheck also displays array statistics such as individual drive information, SuperSwap fan/temperature information, and a plethora of cache statistics such as read hits, write hits, dirty usage, and the number of I/O requests made to the drives themselves (i.e., cache misses). SuperCheck allows remote monitoring of all SuperTrak arrays as long as the machine housing the array is accessible from the remote location (either directly, or via another machine running the Message Server utility). The utility also includes an email-alert notification feature that sends a message to your email address if a drive and/or array connected to the SuperTrak fails. This feature is very common on SCSI RAID cards; few deny its usefulness. Finally, the SuperCheck facilitates array synchronization. This feature compares two mirrored drives sector by sector to ensure that they are identical. If they aren't, data from the primary drive is automatically copied to the secondary drive. The user may also schedule synchronizations for a later time.
Although Promise commands the majority of the ATA RAID market, several other companies realize the industry's potential. 3ware is such a company- in April of 1999, they introduced their first series of ATA RAID controllers: the DiskSwitch 4 series. Soon afterwards, 3ware introduced the Escalade 5000 series in 2, 4, and 8-channel configurations (model numbers 5200, 5400, and 5800, respectively). Contemporary ATA drives were rapidly approching the limit on ATA-33, outdating the 5000 series. To remedy the situation, 3ware released the Escalade 6000 series last summer. Featuring the same 2, 4, and 8-channel options, the 6000 series differs mainly through its ATA-66 support. Like the SuperTrak, the Escalade series has a dedicated ATA channel for each drive; therefore, its ATA-66 interface should not be a limiting factor in the near future. Escalade 6400 Specifications:
The Card... Like the SuperTrak 100, the Escalade 6400 is a full-length PCI card. It features four ATA connectors, two ATA-66 ASICs, and a dedicated RISC processor. The controller features four green LEDs - one next to each ATA connector. They individually light up whenever data is being transferred through each of their respective channels. These LEDs are a welcome feature to folks like us who always want to know as much as possible about what's going on with our hardware at any given time. The Escalade includes the following items:
RAID 5 support for the Escalade series? On February 1st of this year, 3ware introduced a firmware upgrade that brings RAID 5 functionality to the Escalade 6400 and 6800 (not the 6200, of course, since RAID 5 requires at least three drives). Until this point, the Escalade had only supported RAID 0, 1, and 10. How can a firmware upgrade enable a card to support RAID 5? There are two possibilities: 1) The card was always meant to support RAID 5 yet had it delayed by half a year or 2) The card was not meant to support RAID 5, but a decision was made sometime after the its release (perhaps due to the SuperTrak's RAID 5 support?) to implement RAID 5 at the firmware level. More on this later... DiskSwitch and TwinStor: Quality engineering, or fancy marketing? When one visits 3ware's website, it's virtually impossible to miss the words "DiskSwitch" and "TwinStor." These terms have been trademarked by 3ware, and refer to architectural features of their Escalade series of cards. Both terms are covered in great detail on 3ware's site, so we'll just present an overview here. Those seeking more detail may find the respective whitepapers here and here. DiskSwitch The term DiskSwitch actually comprises three aspects of the Escalade controllers: AccelerATA data channels, a packet switching controller, and the card's on-board RISC processor. AccelerATA is 3ware's term for the dedicated ATA channel that each drive enjoys on an Escalade. These channels, along with the packet switching controller and on-board RISC processor, provide each ATA drive with full bandwidth to the host. 3ware claims to be the first company to use packet switching on a storage controller to both lower latency and provide exceptional scaling as more drives are added. TwinStor 3ware's TwinStor architecture - used in RAID 1 and RAID 10 arrays - profiles each drive to allow for "maximum performance for the particular brand of drive used." TwinStor then maintains a statistical history of data accesses to allow sequential and random reads to be distinguished from one other. For random reads, TwinStor reorders I/O's and load balances them between all drives in the array; for sequential I/O's, TwinStor utilizes all drives in the array for maximum transfer rates. Since each drive in a RAID 1 or RAID 10 array has a mirror, proper load balancing of random I/O can provide performance which rivals RAID 0 arrays. The Software... 3ware's Escalade cards include a BIOS that allows arrays to be created or deleted. The BIOS also permits rebuilds and features a toggle for the write cache of an array. (note: Escalade cards do not feature onboard cache; the write cache referred to in BIOS is actually the buffer of the drives themselves). The BIOS also sports a verify feature similar to Promise's synchronization procedure. Just like Promise's SuperCheck, 3ware's software utility, 3DM, allows arrays to be viewed, created, deleted, and maintained. Drives may be added, removed, or designated as hot spares. Also like the SuperCheck utility, 3DM can run in the background as a daemon and provide event notification functionality. The utility can either send an email or trigger a local event in Windows when an array becomes degraded or non-functional. In addition, the card's audible alarm may be toggled and the rebuild rate may be changed (faster rebuilding means lower system performance during rebuilds, and vice-versa). The utility may be accessed via a web browser from any remote location as long as the two machines can connect to each other over a network. All of the utility's aforementioned features are accessible remotely - quite convenient. A Note on Hot Swapping... Though both cards claim hot swap capability, there's a difference between the two. As mentioned above, the SuperTrak Pro comes with three SuperSwap hot swap ATA drive enclosures. These enclosures allow drives to be inserted or removed without having to power down the machine. The Escalade, however, comes with no such enclosure, so a user wishing to take advantage of its hot swap support must purchase third-party hot swap drive enclosures.
StorageReview's RAID reviews will feature WinBench and IOMeter performance measures. The WinBench testing methodology will be identical to that used in the Same Drives - Same Performance? article. To quote from the article: "The following WinBench tests were run five times on each drive: Disk/Read Transfer Rate, Disk CPU utilization, Disk Access Time, Business Disk WinMark 99, and High-End Disk WinMark 99. A single run of each of the above tests was considered a "trial", with five trials being conducted for each drive. The machine was rebooted between trials. Each test's final score represents the average of the five runs." IOMeter, on the other hand, isn't quite as simple. When considered as a whole, StorageReview's three IOMeter access patterns (Workstation, File Server, and Database) are heavily biased towards reads, about 82% to 18%. Though we believe each respective test's read/write distribution represents the tasks from which these patterns draw their names, we realize that write performance becomes a particularly important issue with RAID arrays - especially with RAID levels that use parity. With this in mind, two additional IOMeter access patterns were created to stress write performance. One stresses random writes; the other sequential writes. These new access patterns, as well as the original three, are outlined below.
Information on the testbed may also be found in the aforementioned Same Drives - Same Performance? article. As was the case there, four Maxtor DiamondMax 80 drives will be used for all ATA RAID testing. The Benchmarks... Because the Escalade 6400 "only" supports RAID levels 0, 1, 10, and 5, we can obviously compare it to the SuperTrak only at these four RAID levels. SuperTrak RAID 3-4 benchmarks will follow these comparisons. The SuperTrak's read and write cache were both enabled. The Escalade's "write cache" setting was also enabled. Additionally, all benchmarks in this article were drawn with a stripe size of 64k (save for the SuperTrak RAID 3 tests - see below). We plan on exploring the effect of both stripe size and caching on RAID performance in a future article. All SuperTrak tests were conducted with BIOS revision 1.00 (Build 11) and driver version 1.0.0.0. The Escalade RAID 0/1/10 tests were performed with BIOS revision 1.04.00.009, firmware revision 1.00.43.003, and driver version 1.09.00.005. The firmware was then updated to revision 1.01.18.001 (BIOS revision 1.06.00.009), and driver revision 1.09.000.015 was installed to enable RAID 5 functionality. Since the new firmware and drivers hit the web only days before this review's publish date (3ware's press release still claims an availability date of February 15), and since a trial of IOMeter tests with the new firmware and drivers at RAID levels 0/1/10 revealed no significant performance differences, we chose not to delay this review further by re-running the RAID 0/1/10 tests under the new firmware.
Let's start out with some base scores utilizing a single drive... this gives us something to compare RAID results against. The table below presents WinBench scores for a single DiamondMax 80 on the following controllers: the Abit SL6's on-board ATA controller, a Promise Ultra66, the SuperTrak-100, and the Escalade 6400:
With a single drive, both the SuperTrak and Escalade turn in a Business Disk Winmark score lower than that of the Ultra66. The SuperTrak's High-End score is also significantly lower. These scores seem reasonable for the SuperTrak since its sustained transfer rate is limited to about 20MB/sec when read cache is enabled. For the Escalade however, there's logical reason why its base Business score would be lower than that of the Ultra66 - its sustained transfer rate is the same, and WinBench claims its base seek time is lower. Why is the SuperTrak limited to a sustained transfer rate of just 20MB/sec? Promise tells us it's due to both the cache and the ASICs themselves. Caches in general are bad for sustained transfer rate, Promise says, and the SuperTrak's ASICs were not optimized for STR. This limit will undoubtedly be a disappointment to anyone requiring high STR for optimal performance.
RAID 0 scores from both controllers are rather unimpressive. Generally speaking, each card's Disk Winmark scores are no better than that of a single drive. In the STR arena, the SuperTrak's 20MB/sec causes it to pale in comparison with what the Escalade achieves as more drives are added. As shown above, the Escalade manages an amazing 103MB/sec with a four drive RAID 0 array. Had we not seen it with our own eyes, we probably wouldn't believe such a score: We simply didn't think that the PCI bus's overhead would allow for anything over 90MB/sec or so. It's also interesting to note that average access times for both cards seem to increase somewhat in RAID 0. This phenomenon is fairly consistent in RAID 0 configs, though we're not sure why. Indeed, RAID 0 should yield some positioning benefit in addition to increases in STR. Finally, note the ridiculously high CPU utilization scores for both controllers was in a four drive RAID 0 array. We believe this to be a quirk in WinBench because, as we'll see, IOMeter shows no such CPU utilization increase.
As was the case with RAID 0 scores, the above RAID 1 Disk Winmark scores don't seem to reflect reality. Despite the fact that each card's access time in RAID 1 is significantly lower than that recorded using single drive, and despite the fact that STR is either much better (Escalade) or equal (SuperTrak) to a single drive, RAID 1 Disk Winmark scores are significantly worse. We have a very, very difficult time believing that these scores represent actual performance. As mentioned above, average access time decreases significantly under RAID 1 for both cards. This indicates that both cards perform some type of intelligent load balancing between the two drives. For example, when an I/O request is made to a RAID 1 array, there are always two drives available to service the request since each drive has the exact same data. Therefore, an intelligent RAID card can judge which drive's actuator is closer to the needed data and direct the request accordingly. This tends to result in lower average access times and better performance. The Escalade's STR of 50MB/sec is about 2/3 higher than that of a single drive due to 3ware's TwinStor architecture. Needless to say, it's a significant improvement. Note: Although there are fault tolerance-related differences between RAID 01 and RAID 10, there isn't a theoretical reason for performance differences between these two array levels. Therefore, we feel it's fair to compare the SuperTrak's RAID 01 performance to the Escalade's RAID 10 performance.
Again we face with seemingly illogical Disk Winmark scores. There's no reason why Disk Winmark scores should be so poor here, especially given the fact that both cards enjoy significantly decreased access times in RAID 01/10 (relative to a single drive). In addition, the Escalade's STR doubles in RAID 10; even so, Disk Winmark scores don't appear to reflect this.
Much to our dismay, horrendous WinBench results continued under RAID 5. Note the Disk WinMark scores for both cards: they are not typos. With the SuperTrak, the benchmark would consistently "freeze" 39% through the High-End test for half an hour or more, only to then start up again, and eventually finish. It's certainly not typical behavior... and it's certainly something that can botch up a timed test. This behavior was repeatable on both the testbed and on another machine; we simply don't know if it's an issue with WinBench itself, or the SuperTrak. Regardless, the results are obviously not representative of performance. With the Escalade, there was no "freeze" during the High-End test. However, the results are nonetheless obviously unrepresentative. Once again, the inaccurate results were repeatable on a separate system. Aside from the problems with Disk Winmark scores, it's worth noting that the SuperTrak's RAID 5 access time is higher than that of a single drive configuration as well as any other supported RAID level. This is disappointing, as the SuperTrak is marketed as a RAID 5 controller. The SuperTrak's sustained transfer rate graph is somewhat odd, but we're not sure why. The overall appearance of the graph is very consistent- the dips and peaks in STR always occur in about the same locations. The Escalade's graph is much more "normal", however.
Considering the irregular results that WinBench delivered, IOMeter is a breath of fresh air... it helps us ascertain the true performance of both cards. We're pleased to report that we have no doubt the results below are representative of performance - Promise confirmed that all of the SuperTrak's IOMeter scores are on the mark. We're confident in IOMeter's ability to deliver accurate Escalade results as well. Base scores...
The Escalade's base, single-drive scores are more or less the same as the Ultra66. The SuperTrak, however, lags somewhat here.
Click here to jump to analysis text...
| |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
