 |
|
|---|
The Escalade 7000 Series Revisited: 3Ware's Escalade 7450
|
Before its introduction, 3Ware's Escalade 7000 series was quite possibly the most anticipated ATA RAID card of all time. This was hardly surprising given the then-unmatched performance of the Escalade 6000 series - performance oriented users couldn't wait to see what a successor would bring to the table. In particular, high hopes of improved RAID 5 performance drew the card a lot of interest. Along came the Escalade 7000 series, followed shortly by our review of the 4-port 7410. While the 7410 delivered a huge sequential write performance increase relative to the 6400 in RAID 5, overall RAID 5 scores (i.e., its scores in our File Server, Workstation, and Database tests) were only 2-3% better on average - still significantly lower than those of the Adaptec 2400A, and thus a disappointment to many. But when it comes to improving RAID 5 performance, 3Ware certainly wasn't about to give up, as is evidenced by the introduction of the Escalade 7x50 series. The Escalade 7450... Compared to the Escalade 7x10 series, the Escalade 7x50 series is about one thing and one thing only: improved RAID 5 performance. The hardware itself, however, is virtually identical to that of the 7x10 series - the difference is in firmware. Dubbed R5 Fusion Technology by the marketing folks, this new firmware aims to significantly improve RAID 5 write performance, and hence, overall RAID 5 performance as well. Does R5 Fusion deliver? We'll find out shortly, but first, a quick look at the card's specs (which are, again, pretty much identical to those of the 7410):
Included with the card are the following items:
The Software... The Escalade 7450 remains unchanged from the 7410 as far as software goes. See our review of the 7410 for a rundown of the software that accompanies each card. The Benchmarks... The following benchmarks were run with the card's write cache enabled and a stripe size of 64k. Driver release 7.3 (version 1.10.01.036) was used along with BIOS version 1.07.01.17.
Base Scores...
Nothing too exciting here... The Escalade 7450's access time is a bit lower than that of the 7410, but nothing may be concluded from this test alone. RAID 0...
3Ware continues to find new ways to push the maximum realizable bandwidth of the standard 32-bit/33MHz PCI bus to levels that most (including myself) never thought possible. The Escalade 7450's score of 116MB/sec is 9MB/sec higher than that of the 7410, and 6MB/sec higher than the second-place ATA RAID 2400A. RAID 1...
The 7450's RAID 1 WinBench scores are, unsurprisingly, virtually identical to the 7410's. RAID 10...
The 7450's RAID 10 access time in WinBench returns to Escalade 6400 territory - about 1.7ms lower than that of the 7410. In the Escalade 7410 review, however, IOMeter seemed to disagree with this disparity between the 7410 and 6400 (i.e., scores under linear load were very similar), so we're not sure what (if anything) to make of this. RAID 5...
Scores between the 7450 and 7410 again remain very similar. Is this an indication that R5 Fusion Technology is nothing but marketing? Seems unlikely - it would be foolish to draw any conclusions based solely on a card's sustained transfer rate, linear access time, and CPU usage during 4MB/sec sustained reads (yep, that's how WinBench measures CPU usage by default). Performance under various load levels and access patterns is what matters, so we'll look to IOMeter to give us a more detailed performance picture.
Base Scores...
The 7450 turns in a score very similar to the 7410 with a single drive. RAID 0...
Scores between the 7450 and 7410 are again very similar. The largest difference is the 7450's 1.5MB/sec lead in sequential writes in three and four drive arrays. RAID 1...
Again, scores are virtually the same between the 7450 and 7410. Unfortunately, Sequential Write CPU usage also remains the same... Readers of the 7410 review may recall that the adapter had an apparent driver issue that caused CPU usage to top out at 100% in the Sequential Write test in RAID levels 1, 10, and 5. The 7450 also exhibits this behavior in RAID 1:
UPDATE: After the initial benchmarks, 3ware sent us updated drivers (1.10.01.042) for the 7x50 series meant to address the high CPU usage issue seen above. The table below compares the 7450's RAID 1 Sequential Write CPU usage with the two drivers:
The new drivers decrease CPU utilization by about 58% under moderate and heavy load - CPU usage under linear, very light, and light loads is unaffected, however. Despite the improvement, the numbers are still much higher than is desirable. Hopefully, a future driver release will lower them even further. RAID 10...
The 7450 scores (surprise surprise) about the same in RAID 10 as the 7410. This goes for CPU usage as well:
UPDATE: Below is the CPU usage comparison table for RAID 10:
As with RAID 1, CPU usage is improved with the new drivers, but not as much as we'd all like. At this point, some readers may be asking themselves why 3Ware even bothered to release a card that performs exactly the same as its older brother. Keep in mind what we said earlier: the 7450 is aimed at improving RAID 5 performance. Let's have a look at its RAID 5 scores! RAID 5...
It's here that significant performance differences arise between the 7450 and 7410. In the File Server, Workstation, and Database access patterns, the 7450 consistently scores 5-15% better than the 7410. The Database pattern enjoys the greatest improvement out of the three - likely because it incorporates the greatest percentage of writes. Looking at Random Write scores, one sees that the 7450 scores significantly better than the 7410 - especially under light loads. The largest improvement, however, arrives in Sequential Write tests. The 7410's average score of 25MB/sec was double that of the competition in its own right. The 7450, however, weighs in at an unprecedented 45MB/sec with a 4-drive array - a very impressive performance boost, especially considering that it comes from firmware alone. Adding to the good news is the fact that the aforementioned Sequential Write CPU usage issue isn't present under RAID 5 as it was with the 7410:
After our Escalade 7410 review, several readers wrote in asking why tests weren't conducted in a 64-bit PCI slot. They were concerned that the card's performance was limited by the testbed's 32-bit/33MHz PCI bus, and that we weren't giving it a fair shake by not testing it in a 64-bit slot. Our response to these concerns is as follows: the only thing that 64-bit PCI offers relative to 32-bit PCI is extra bandwidth. Since the vast majority of our tests stress random IO performance rather than sequential IO performance, and since the random IO performance of current drives is in no way limited by the bandwidth of a standard PCI bus (rather by access time), 64-bit PCI would make no difference in these tests. Unfortunately, we couldn't test our theory at the time because we didn't have any 64-bit PCI capable motherboards. However, thanks to 3Ware, the time has come to see how much of a difference 64-bit PCI makes in our tests. Wanting to see the 7450 tested in a 64-bit slot as well as a 32-bit slot, 3Ware sent us the following hardware:
The 370DE6, along with the processors and RAM, obviously constitutes a high-end system. We were very enthusiastic about seeing how the Escalade 7450 would fare in such a system compared to our standard testbed. To prep the system for testing, Win2000 was installed on one of our Cheetah 18XL's. The process used for this installation was the same as that used for the standard testbed back in January. The 7450 was tested in the first PCI slot. No other PCI devices were present in the system. Since WinBench Diskmark scores are heavily affected by system components such as processors and RAM, we chose not to run these tests in the SuperMicro system. The only WinBench test run was the sustained transfer rate test. Our IOMeter test suite was, of course, run in its entirety. The Results... Base Scores...
With a single drive, there are no significant differences between the two systems - the 370DE6 scores about 1% better all around. RAID 0...
With the Escalade 7450 achieving 116MB/sec of STR performance in the testbed, there wasn't much room for improvement in the 370DE6 system. Each Diamondmax 80 used for testing can sustain just under 30MB/sec - four DM80's, therefore, can't do much better than 116MB/sec. Indeed, the score of 118MB/sec achieved in the 370DE6 system is clearly a limitation of the drives themselves as opposed to the 7450. In the File Server, Workstation, Database, and Random Write tests, scores remain virtually the same between the two systems - even as performance scales along with the number of drives in the array. Sequential write performance, however, is another matter... The Escalade 7450 is limited to around 67MB/sec of write performance when connected to the testbed's 32-bit/33MHz PCI bus. In the 370DE6 system, however, there is no such limitation - the 7450's sequential write performance scales quite nicely to 103MB/sec. The obvious question: why is the Escalade 7450 limited to 67MB/sec of sequential write performance in the testbed when its sequential read performance in that very same machine is 116MB/sec? It's impossible to know what the exact cause is here... It may be that the 7450 is, for whatever reason, much less bandwidth-efficient when performing sequential writes relative to sequential reads. It may also be that the realizable bandwidth of the PCI bus itself is much lower for writes than it is for reads. Whatever the reason, 64-bit PCI is clearly the way to go for maximum sequential write performance with the Escalade 7450. RAID 1...
In the File Server, Workstation, Database, and Random Write patterns, scores are again virtually the same, with the 370DE6 system leading by a very slim 1% or so overall. Its low Sequential Write scores are of interest, though... There appears to be an issue with the 7450 in this particular machine. The sequential write rate would commence at the expected 26-27MB/sec at the beginning of each subtest, only to drop into the 12MB/sec range a few minutes in. We re-ran this test several times and received the same results. RAID 10...
Scores again remain practically identical save for the Sequential Write test - the 370DE6 system wins big time here, just like it did in RAID 0. RAID 5...
The story is once again the same... The 7450's sequential write performance clocks in significantly higher in the 370DE6 system, while all other scores are identical.
The Escalade 7x50 series was designed to do one thing, and one thing only: improve RAID 5 performance relative to the 7x10 series. Overall, it comes through pretty nicely... Although it's overall RAID 5 performance can't quite match that of Adaptec's ATA RAID 2400A (in 3-drive RAID 5, the 7450 trails by 10-15% on average in the File Server, Workstation, and Database tests, a difference reduced to about 5% when adding a 4th drive to the mix), its RAID 5 sequential write performance (45MB/sec in the testbed; 60MB/sec in the 370DE6 system) is unheard of - the Escalade 7x50 is obviously the card of choice for anyone who demands top sequential write performance in RAID 5. It's amazing what a little firmware magic can do: the 7450's 5-15% RAID 5 performance increase relative to the 7410 in the File Server, Workstation, Database, and Random Write tests, along with its remarkably improved RAID 5 sequential write performance, was all accomplished via different firmware, not different hardware. This demonstrates that firmware can be just as important as the hardware itself. Concerns about performance differences between 32-bit and 64-bit PCI have hopefully been put to bed. Although the 7450's sequential write performance is indeed significantly better in a 64-bit PCI slot, the fact remains that its performance in our core, random IO tests (File Server, Workstation, and Database) remains unchanged by 64-bit PCI. This shouldn't be surprising - relative to 32-bit PCI, 64-bit PCI simply brings more bandwidth to the table. Random IO performance, however, isn't bottlenecked by bandwidth - average access time is the limitation here. | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Due to the consistently unrepresentative WinBench Diskmark scores we've been getting in our RAID reviews, we're no longer going to bother to comment on them in RAID articles. We'll continue to present them for the sake of completeness, but the only WinBench tests that we'll discuss are the sustained transfer rate, access time, and CPU utilization tests.
HOME | ARTICLES | LEADERBOARD | PERFORMANCE DATABASE | REFERENCE GUIDE
COMMUNITY | RELIABILITY SURVEY | SUPPORT SR! | ABOUT SR |
Copyright © 1998-2005 StorageReview.com, Inc. All rights reserved.
Write: Webmaster