Home |
Search |
Today's Posts |
#1
|
|||
|
|||
100 base T network capacity
I want to set up a network music server with one client, a single machine in
my listening room. Since actual throughput on a 100 mb network depends to a great extent on factors other than the actual physical layer, I'd appreciate any rules of thumb determined by experience. A single channel is about 2.3 mb/second, 8 channels = 18.4 mb/second. I want to feed an 8-channel D/A, at 96/24 resolution, directly from Adobe Audition, but the files will be on the server. Both machines will run Windows 2000, with Ethernet adaptors on standard 32 bit PCI buses. Is this actually practical with desktop machines you've used? |
#2
|
|||
|
|||
"Robert Morein" wrote in message ...
I want to set up a network music server with one client, a single machine in my listening room. Since actual throughput on a 100 mb network depends to a great extent on factors other than the actual physical layer, I'd appreciate any rules of thumb determined by experience. A single channel is about 2.3 mb/second, 8 channels = 18.4 mb/second. I want to feed an 8-channel D/A, at 96/24 resolution, directly from Adobe Audition, but the files will be on the server. Both machines will run Windows 2000, with Ethernet adaptors on standard 32 bit PCI buses. Is this actually practical with desktop machines you've used? Doing the math 100 Mb should handle the load. However, considering the low cost of network adapters I would recommend going Gigabit anyway. |
#3
|
|||
|
|||
"Robert Morein" wrote in message
I want to set up a network music server with one client, a single machine in my listening room. Since actual throughput on a 100 mb network depends to a great extent on factors other than the actual physical layer, I'd appreciate any rules of thumb determined by experience. A single channel is about 2.3 mb/second, 8 channels = 18.4 mb/second. I want to feed an 8-channel D/A, at 96/24 resolution, directly from Adobe Audition, but the files will be on the server. Both machines will run Windows 2000, with Ethernet adaptors on standard 32 bit PCI buses. Is this actually practical with desktop machines you've used? What you're doing relates to me recording and mixing 12 32/44 channels which I do reliably all the time, except you're talking about doubling the load. The simple answer is bag the 96 KHz sampling, but I know your audio politics. 100BTX is not the most efficient user of media around. Bursts of 20 MB/Sec are doable, but for this application we need to look at the whole system. If you have a disk subsystem that is capable of a sustained 20 MB/sec locally, its not going to be 20 MB/Sec across a 100BTX network. Based on my experience, getting a reliable, sustained 20 MB/sec off a commodity 7200 rpm hard drive *without* a network layer in place can be iffy unless the environment is pretty well controlled and the disks have pretty high performance. For example, I would at minimum recommend a l hard drive subsystem composed something like a well-defragged 2-way stripe set made up of 7200 rpm 160 GB ATA drives. You don't need the striping for speed when the array is nearly empty, but you probably will need striping when it is nearly full. Striping gets you consistancy, not raw speed. |
#4
|
|||
|
|||
"Robert Morein" wrote in
: Since actual throughput on a 100 mb network depends to a great extent on factors other than the actual physical layer, I'd appreciate any rules of thumb determined by experience. A single channel is about 2.3 mb/second, 8 channels = 18.4 mb/second. Are these MegaBytes per second or Megabits? Because a 100 base T network runs at a theoretical 100megabits or ~12.8MB/s. In reality you're looking at 80mbits top or 10MB/s. I want to feed an 8-channel D/A, at 96/24 resolution, directly from Adobe Audition, but the files will be on the server. Both machines will run Windows 2000, with Ethernet adaptors on standard 32 bit PCI buses. Is this actually practical with desktop machines you've used? You may want to upgrade to Gigabit ethernet. Someone else mentioned hard drive speed - modern hard drives can easily sustain 30MB/s if the drive is not fragmented heavily. To ensure reliability you might want to put a dedicated recording disk on the server. Drives are cheap - you can pick up a very good 80GB disk for ~ 60.00USD. The striped RAID idea is good too, if you have enough drive bays. But just be aware... on a striped array, if one of the 2 disks crash, you lose ALL the data on both drives. -- Lucas Tam ) Please delete "REMOVE" from the e-mail address when replying. http://members.ebay.com/aboutme/coolspot18/ |
#5
|
|||
|
|||
Lucas Tam wrote:
"Robert Morein" wrote in : Since actual throughput on a 100 mb network depends to a great extent on factors other than the actual physical layer, I'd appreciate any rules of thumb determined by experience. A single channel is about 2.3 mb/second, 8 channels = 18.4 mb/second. Are these MegaBytes per second or Megabits? Because a 100 base T network runs at a theoretical 100megabits or ~12.8MB/s. In reality you're looking at 80mbits top or 10MB/s. I'm pretty sure he means megabits/second. The calculation he's using is probably this: 96000 samples/second * 24 bits/sample = 2,304,000 bits/second and 2,304,000 bits/second/channel * 8 channels = 18,432,000 bits/second So, those numbers would be 288 kbytes/second and 2304 kbytes/second, respectively. In which case, this is well within the bandwidth that 100BaseT can handle. The only question is whether the actual hardware and software involved is able to sustain an average of 2.3 megabytes/second. Most fast PCs these days should be able to do that. Of course, it really depends on the software involved and also whether the network cards are cheap pieces of junk or something reasonable. - Logan |
#6
|
|||
|
|||
Logan Shaw wrote in
: In which case, this is well within the bandwidth that 100BaseT can handle. The only question is whether the actual hardware and software involved is able to sustain an average of 2.3 megabytes/second. Most fast PCs these days should be able to do that. Yes definately, most modern harddrives made within the last 2 years can sustain between 30MB/s - 60MB/s depending on the location of the data. Of course, it really depends on the software involved and also whether the network cards are cheap pieces of junk or something reasonable. Even an el cheapo network card should be able to do 70 - 80mbit/s no problem : ) -- Lucas Tam ) Please delete "REMOVE" from the e-mail address when replying. http://members.ebay.com/aboutme/coolspot18/ |
#7
|
|||
|
|||
"Arny Krueger" wrote in message ... "Robert Morein" wrote in message I want to set up a network music server with one client, a single machine in my listening room. Since actual throughput on a 100 mb network depends to a great extent on factors other than the actual physical layer, I'd appreciate any rules of thumb determined by experience. A single channel is about 2.3 mb/second, 8 channels = 18.4 mb/second. I want to feed an 8-channel D/A, at 96/24 resolution, directly from Adobe Audition, but the files will be on the server. Both machines will run Windows 2000, with Ethernet adaptors on standard 32 bit PCI buses. Is this actually practical with desktop machines you've used? What you're doing relates to me recording and mixing 12 32/44 channels which I do reliably all the time, except you're talking about doubling the load. The simple answer is bag the 96 KHz sampling, but I know your audio politics. 100BTX is not the most efficient user of media around. Bursts of 20 MB/Sec are doable, but for this application we need to look at the whole system. If you have a disk subsystem that is capable of a sustained 20 MB/sec locally, its not going to be 20 MB/Sec across a 100BTX network. Based on my experience, getting a reliable, sustained 20 MB/sec off a commodity 7200 rpm hard drive *without* a network layer in place can be iffy unless the environment is pretty well controlled and the disks have pretty high performance. For example, I would at minimum recommend a l hard drive subsystem composed something like a well-defragged 2-way stripe set made up of 7200 rpm 160 GB ATA drives. You don't need the striping for speed when the array is nearly empty, but you probably will need striping when it is nearly full. Striping gets you consistancy, not raw speed. Arny, thanks for your constructive answer. As some of the other posters have commented, there may be a confusion between bits and bytes. Here are some measurements I've made using a DVD burning program, a similar, but not identical problem: Via P4X266 chipset with 2.6g P4: 5.2 megabytes/second streaming to a DVD burner = 40 megabits/second Via 266 Socket 370 chipset with Via C3 Nehemiah : exactly the same. Tyan S2460 dual MP chipset: 14.4 megabytes/second = 115 megabits/second In each case, a single 7200 rpm disk is the source. Raw speed for this type of device is approximately 500 megabits/second. This leads me to think that striping is not required. The question is, how much will the network interface slow it down? |
#8
|
|||
|
|||
On 12/2/2004 6:39 PM, Robert Morein wrote:
"Arny Krueger" wrote in message ... "Robert Morein" wrote in message I want to set up a network music server with one client, a single machine in my listening room. Since actual throughput on a 100 mb network depends to a great extent on factors other than the actual physical layer, I'd appreciate any rules of thumb determined by experience. A single channel is about 2.3 mb/second, 8 channels = 18.4 mb/second. I want to feed an 8-channel D/A, at 96/24 resolution, directly from Adobe Audition, but the files will be on the server. Both machines will run Windows 2000, with Ethernet adaptors on standard 32 bit PCI buses. Is this actually practical with desktop machines you've used? What you're doing relates to me recording and mixing 12 32/44 channels which I do reliably all the time, except you're talking about doubling the load. The simple answer is bag the 96 KHz sampling, but I know your audio politics. 100BTX is not the most efficient user of media around. Bursts of 20 MB/Sec are doable, but for this application we need to look at the whole system. If you have a disk subsystem that is capable of a sustained 20 MB/sec locally, its not going to be 20 MB/Sec across a 100BTX network. Based on my experience, getting a reliable, sustained 20 MB/sec off a commodity 7200 rpm hard drive *without* a network layer in place can be iffy unless the environment is pretty well controlled and the disks have pretty high performance. For example, I would at minimum recommend a l hard drive subsystem composed something like a well-defragged 2-way stripe set made up of 7200 rpm 160 GB ATA drives. You don't need the striping for speed when the array is nearly empty, but you probably will need striping when it is nearly full. Striping gets you consistancy, not raw speed. Arny, thanks for your constructive answer. As some of the other posters have commented, there may be a confusion between bits and bytes. Here are some measurements I've made using a DVD burning program, a similar, but not identical problem: Via P4X266 chipset with 2.6g P4: 5.2 megabytes/second streaming to a DVD burner = 40 megabits/second Via 266 Socket 370 chipset with Via C3 Nehemiah : exactly the same. Tyan S2460 dual MP chipset: 14.4 megabytes/second = 115 megabits/second In each case, a single 7200 rpm disk is the source. Raw speed for this type of device is approximately 500 megabits/second. This leads me to think that striping is not required. The question is, how much will the network interface slow it down? Make sure the network is switched, not hub. Get RAID 1 of 2 SATA drives or Ultra 320 SCSI and 1 GB of DDR RAM and Windows XP in the music server. Make sure FULL DUPLEX is enabled for all computers on LAN. |
#9
|
|||
|
|||
"Robert Morein" wrote in
: In each case, a single 7200 rpm disk is the source. Raw speed for this type of device is approximately 500 megabits/second. This leads me to think that striping is not required. The question is, how much will the network interface slow it down? The disk is much faster than the network interface, so the interface will run at full speed - which is usually ~60 - 80% of capacity, depend on the number of machines on the network. -- Lucas Tam ) Please delete "REMOVE" from the e-mail address when replying. http://members.ebay.com/aboutme/coolspot18/ |
#10
|
|||
|
|||
"Lucas Tam" wrote in message .. . "Robert Morein" wrote in : In each case, a single 7200 rpm disk is the source. Raw speed for this type of device is approximately 500 megabits/second. This leads me to think that striping is not required. The question is, how much will the network interface slow it down? The disk is much faster than the network interface, so the interface will run at full speed - which is usually ~60 - 80% of capacity, depend on the number of machines on the network. Full speed, but there must be some overhead in handling the packets. I understand some network cards are faster than others, depending upon how much work the driver gives the CPU. Afew network cards do all the low level processing on the card. ??? |
#11
|
|||
|
|||
"Robert Morein" wrote in
: The disk is much faster than the network interface, so the interface will run at full speed - which is usually ~60 - 80% of capacity, depend on the number of machines on the network. Full speed, but there must be some overhead in handling the packets. Yes there are bottlenecks on several levels, MAC level, TCP, and software. Hence that is why the interface will probably run at 60 - 80% of the maximum theoretical speed. You'll never hit 100mbit on ethernet. I understand some network cards are faster than others, depending upon how much work the driver gives the CPU. Afew network cards do all the low level processing on the card. ??? Yes some cards do - mostly server grade network cards. But considering how fast modern computers are... any computer built within the last 3 - 5 years will saturate a 100mbit connection. -- Lucas Tam ) Please delete "REMOVE" from the e-mail address when replying. http://members.ebay.com/aboutme/coolspot18/ |
#12
|
|||
|
|||
I have experienced about 40 megabits per second on a 100 megabit ethernet, 2
computers on the network. Bob "Robert Morein" wrote in message ... I want to set up a network music server with one client, a single machine in my listening room. Since actual throughput on a 100 mb network depends to a great extent on factors other than the actual physical layer, I'd appreciate any rules of thumb determined by experience. A single channel is about 2.3 mb/second, 8 channels = 18.4 mb/second. I want to feed an 8-channel D/A, at 96/24 resolution, directly from Adobe Audition, but the files will be on the server. Both machines will run Windows 2000, with Ethernet adaptors on standard 32 bit PCI buses. Is this actually practical with desktop machines you've used? |
#13
|
|||
|
|||
In article ,
"Robert Morein" wrote: I want to set up a network music server with one client, a single machine in my listening room. Since actual throughput on a 100 mb network depends to a great extent on factors other than the actual physical layer, I'd appreciate any rules of thumb determined by experience. A single channel is about 2.3 mb/second, 8 channels = 18.4 mb/second. I want to feed an 8-channel D/A, at 96/24 resolution, directly from Adobe Audition, but the files will be on the server. Both machines will run Windows 2000, with Ethernet adaptors on standard 32 bit PCI buses. Is this actually practical with desktop machines you've used? You might be better off with Gigabit with jumbo frames. Jumbo frames enlarge the data capacity of a packet from 1500 bytes to 9000 bytes. It cuts down on the CPU overhead needed to manage each packet. The downside is that jumbo frames aren't quite mainstream yet. It's hard to find hardware that handles it properly at a reasonable cost. |
#14
|
|||
|
|||
"Robert Morein" wrote in message
I want to set up a network music server with one client, a single machine in my listening room. Since actual throughput on a 100 mb network depends to a great extent on factors other than the actual physical layer, I'd appreciate any rules of thumb determined by experience. A single channel is about 2.3 mb/second, 8 channels = 18.4 mb/second. Since PC data rates are usually given in bytes per second, I missed the fact that these numbers use the lower case b, and are bits per second. Using standard units, a 24/96 channel moves 96,000 3 byte words per second or 288,000 bytes per second. 8 24/96 audio channels is a modest 2,304,000 bytes per second or 2.3 megabytes per second. 2.3 megabytes per second is a modest load, compared to the real-world sustaned data rates possible with modern 7200 rpm drives in the 100-200 gigabyte range on modern 2 GHz plus CPU-type machines.. It's something like a tenth of their typical data transfer capacity if they aren't badly fragmented. 2.3 megabytes per second is even well within the capacity of a legacy file server, such as a P2-233 running windows NT 4 with 256 megs of RAM and old 6 gigabyte 5400 rpm hard drives on a 100BTX network. That sort of a system can be expected to deliver data at a sustained rate of about 6 megabytes per second. I want to feed an 8-channel D/A, at 96/24 resolution, directly from Adobe Audition, but the files will be on the server. Both machines will run Windows 2000, with Ethernet adaptors on standard 32 bit PCI buses. Is this actually practical with desktop machines you've used? With modern equipment, it is a cakewalk. Been there done that routinely. As others have pointed out, the most serious bottleneck in a modern system is the 12.5 megabyte per second absolute maximum theoretical data rate of the core of a 100BTX network. In the real world only about 80 or 90 percent of that is available. Last time I moved bulk data over a 100BTX network data sustained data rates were on the order of 8 megs per second as observed with a stopwatch. That was a disk-disk transfer over a network with about a half dozen machines on it. |
#15
|
|||
|
|||
Arny Krueger wrote: "Robert Morein" wrote in message I want to set up a network music server with one client, a single machine in my listening room. Since actual throughput on a 100 mb network depends to a great extent on factors other than the actual physical layer, I'd appreciate any rules of thumb determined by experience. A single channel is about 2.3 mb/second, 8 channels = 18.4 mb/second. Since PC data rates are usually given in bytes per second, I missed the fact that these numbers use the lower case b, and are bits per second. Using standard units, a 24/96 channel moves 96,000 3 byte words per second or 288,000 bytes per second. It's not 3 byte but 4 byte. No CPU handles 24bit, but 32bit. The sampled 24bit are either left-aligned or right-aligned 32bit integers. Concerning the network speed: The error correction might take another 10% of speed. I would guess a 50% of 5-6 MB/s is realistic in case of 100 mb network. Kind Regards Nudge -- Nudge // PCS Records Studio Leipzig http://studio.lieber-media.de |
#16
|
|||
|
|||
Your actual throughput would probably be around 5 megabytes per second.
Bob "Lucas Tam" wrote in message .. . "Robert Morein" wrote in : Since actual throughput on a 100 mb network depends to a great extent on factors other than the actual physical layer, I'd appreciate any rules of thumb determined by experience. A single channel is about 2.3 mb/second, 8 channels = 18.4 mb/second. Are these MegaBytes per second or Megabits? Because a 100 base T network runs at a theoretical 100megabits or ~12.8MB/s. In reality you're looking at 80mbits top or 10MB/s. I want to feed an 8-channel D/A, at 96/24 resolution, directly from Adobe Audition, but the files will be on the server. Both machines will run Windows 2000, with Ethernet adaptors on standard 32 bit PCI buses. Is this actually practical with desktop machines you've used? You may want to upgrade to Gigabit ethernet. Someone else mentioned hard drive speed - modern hard drives can easily sustain 30MB/s if the drive is not fragmented heavily. To ensure reliability you might want to put a dedicated recording disk on the server. Drives are cheap - you can pick up a very good 80GB disk for ~ 60.00USD. The striped RAID idea is good too, if you have enough drive bays. But just be aware... on a striped array, if one of the 2 disks crash, you lose ALL the data on both drives. -- Lucas Tam ) Please delete "REMOVE" from the e-mail address when replying. http://members.ebay.com/aboutme/coolspot18/ |
#17
|
|||
|
|||
"Robert Morein" wrote in message
"Arny Krueger" wrote in message In each case, a single 7200 rpm disk is the source. Raw speed for this type of device is approximately 500 megabits/second. This leads me to think that striping is not required. Striping is for reliability, not maximum speed. Disks have vastly different speeds as they fill up. Striping addresses that. The question is, how much will the network interface slow it down? I don't think I've ever seen file copies at more than 8 megabytes per second on a 100BTX network. Theoretical max is more like 12 megabytes per second. Of course, other network activity will also impact it. |
#18
|
|||
|
|||
In article , "Arny Krueger"
wrote: In each case, a single 7200 rpm disk is the source. Raw speed for this type of device is approximately 500 megabits/second. This leads me to think that striping is not required. Striping is for reliability, not maximum speed. Disks have vastly different speeds as they fill up. Striping addresses that. You mean of course, that RAID striping "1" is for reliability and RAID striping "0" is for speed, to clarify. In addition, there are variations of these two basic types. |
#19
|
|||
|
|||
jackfish wrote:
In article , "Arny Krueger" wrote: In each case, a single 7200 rpm disk is the source. Raw speed for this type of device is approximately 500 megabits/second. This leads me to think that striping is not required. Striping is for reliability, not maximum speed. Disks have vastly different speeds as they fill up. Striping addresses that. You mean of course, that RAID striping "1" is for reliability and RAID striping "0" is for speed, to clarify. In addition, there are variations of these two basic types. As long as we're clarifying, the "R" in RAID means "redundant". Since striping by itself is not redundant, it is strictly speaking not a form of RAID. That's why people often calling it RAID 0. The official levels of RAID start with RAID 1, so calling it RAID 0 indicates that it's not really a part of the club. Actually, a lot of this stuff is very complex. For large data transfers (such as is done in audio, *if* your drive is not too fragmented) striping helps. But for lots of small random access transfers, mirroring (RAID 1) can actually be faster, but only for reading and not for writing. RAID 5 is an attractive option because it can (if properly implemented) do fast sequential reads and writes[1] and it provides increased reliability in that TWO drives have to fail for you to actually lose any data, and the odds of two drives failing are generally much lower than double the odds of one failing. (But if they do fail, then you've still lost all your data, just like with striping.) - Logan [1] If you issue a write for an entire stripe width, there is no need to read any data from anywhere to recompute new parity data; you can just compute fresh parity data for the whole stripe and overwrite the other data. |
#20
|
|||
|
|||
"jackfish" wrote in message
In article , "Arny Krueger" wrote: In each case, a single 7200 rpm disk is the source. Raw speed for this type of device is approximately 500 megabits/second. This leads me to think that striping is not required. Striping is for reliability, not maximum speed. Disks have vastly different speeds as they fill up. Striping addresses that. You mean of course, that RAID striping "1" is for reliability and RAID striping "0" is for speed, to clarify. In addition, there are variations of these two basic types. Actually, I mean that conventionally, disks striped for speed are striped complementarily. IOW one disc is striped from the outside in, and the other is striped from the inside out. This effectively takes the average DTR of the inner and out tracks and multiplies it by two. Thus, the pair of discs striped in this fashion has more reliable performance characteristics, its speed does not drop nearly as preciptously when it fills up, as does a single disk. Ironically, the outer tracks of a single disk may be faster than the pair of discs striped this way. |
#21
|
|||
|
|||
Arny Krueger wrote:
Actually, I mean that conventionally, disks striped for speed are striped complementarily. IOW one disc is striped from the outside in, and the other is striped from the inside out. This effectively takes the average DTR of the inner and out tracks and multiplies it by two. Yes, when you take the sum of two things and divide it by two and then multiply it again, you get the sum again. :-) :-) Thus, the pair of discs striped in this fashion has more reliable performance characteristics, I think I might call it consistent rather than reliable. But I do see the point. By the way, I have never seen a system that actually does striping the way you describe, even though it is a useful idea. One other concept is to just leave the first 1/3 or so of the capacity of both drives empty. If the filesystem doesn't even include those parts of the disks, there's little chance of them being used when the disk fills up. If you want, use the first 1/3 of both disks for scratch space or archives or backups or something (as long as it's not something you need to access during the time you expect to get good performance out of the stripe set). - Logan |
#22
|
|||
|
|||
"Arny Krueger" wrote in
: "jackfish" wrote in message In article , "Arny Krueger" wrote: In each case, a single 7200 rpm disk is the source. Raw speed for this type of device is approximately 500 megabits/second. This leads me to think that striping is not required. Striping is for reliability, not maximum speed. Disks have vastly different speeds as they fill up. Striping addresses that. You mean of course, that RAID striping "1" is for reliability and RAID striping "0" is for speed, to clarify. In addition, there are variations of these two basic types. Actually, I mean that conventionally, disks striped for speed are striped complementarily. IOW one disc is striped from the outside in, and the other is striped from the inside out. This effectively takes the average DTR of the inner and out tracks and multiplies it by two. Thus, the pair of discs striped in this fashion has more reliable performance characteristics, its speed does not drop nearly as preciptously when it fills up, as does a single disk. Ironically, the outer tracks of a single disk may be faster than the pair of discs striped this way. For reliability and speed the best raid option available is RAID 1+0 sometime called RAID 10. Not all vendors have that available so the next best is RAID5 which means that you can lose one disc and your data is safe. WIth raid 10 you can lose multiple disks and your data will be safe. r -- Nothing beats the bandwidth of a station wagon filled with DLT tapes. |
#23
|
|||
|
|||
"Logan Shaw" wrote in message
Arny Krueger wrote: Actually, I mean that conventionally, disks striped for speed are striped complementarily. IOW one disc is striped from the outside in, and the other is striped from the inside out. This effectively takes the average DTR of the inner and out tracks and multiplies it by two. Yes, when you take the sum of two things and divide it by two and then multiply it again, you get the sum again. :-) :-) Thus, the pair of discs striped in this fashion has more reliable performance characteristics, I think I might call it consistent rather than reliable. But I do see the point. I agree that consistent might be a better choice of words. By the way, I have never seen a system that actually does striping the way you describe, even though it is a useful idea. Interesting. I've never seen one that doesn't do it this way. Vendors don't seem to talk much about this, but it explains why striping a pair of disks can yield a stripe set that runs slower than either disck, when the disks are mostly empty. One other concept is to just leave the first 1/3 or so of the capacity of both drives empty. If the filesystem doesn't even include those parts of the disks, there's little chance of them being used when the disk fills up. If you want, use the first 1/3 of both disks for scratch space or archives or backups or something (as long as it's not something you need to access during the time you expect to get good performance out of the stripe set). My view of this is that it might make sense to partition a disk into multiple regions that inherently vary in terms of perforamnce. Then assign data to the appropriate region based on its performance needs. |
#24
|
|||
|
|||
Logan Shaw wrote in
: Arny Krueger wrote: Actually, I mean that conventionally, disks striped for speed are striped complementarily. IOW one disc is striped from the outside in, and the other is striped from the inside out. This effectively takes the average DTR of the inner and out tracks and multiplies it by two. Yes, when you take the sum of two things and divide it by two and then multiply it again, you get the sum again. :-) :-) Thus, the pair of discs striped in this fashion has more reliable performance characteristics, I think I might call it consistent rather than reliable. But I do see the point. By the way, I have never seen a system that actually does striping the way you describe, even though it is a useful idea. One other concept is to just leave the first 1/3 or so of the capacity of both drives empty. If the filesystem doesn't even include those parts of the disks, there's little chance of them being used when the disk fills up. If you want, use the first 1/3 of both disks for scratch space or archives or backups or something (as long as it's not something you need to access during the time you expect to get good performance out of the stripe set). - Logan I don't know where Arny got the idea that disks are striped as described, but that isn't how it is done in the real world. On a write it would not matter one bit if one disc was writing from the outside in or inside out. The point of striping is to improve speed and improve speed it does dramatically. I have seen large arrays write a gigabyte of data per second and there was none of that business of one disk starting at block 0 and the other starting at EOF. r |
#25
|
|||
|
|||
Arny Krueger wrote:
"Logan Shaw" wrote in message By the way, I have never seen a system that actually does striping the way you describe, even though it is a useful idea. Interesting. I've never seen one that doesn't do it this way. Vendors don't seem to talk much about this, but it explains why striping a pair of disks can yield a stripe set that runs slower than either disck, when the disks are mostly empty. There are a couple of miscellaneous reasons that a striped set can run slower than an individual disk. The first reason is that both disk heads may have to seek to satisfy an I/O. Let's say you want to read a small amount of data, and it stretches across the boundary from one stripe to the next. If you had just one drive, it would seek to the beginning of your data and 99% chance it would read it all in one continuous transfer. But with striped disks, both disks' heads are in the wrong position (probably), and both have to move to the right position and read before you get back the data you want. In most cases, they aren't in the same position to start with, so you have to wait longer for one than the other. Another way of looking at this is that if you do an I/O that is larger than the stripe width (or overlaps a stripe boundary), then you are getting several disks involved and you have to wait around for *all* of them to finish before you can proceed. The second reason is sort of similar to the first, but different. When you want to get some data off a disk, you have to wait for the head to get where it needs to be (near the center or out on the edge), but you also have to wait for the platter to spin around until your data passes under the head. Most striped disk sets do not have the spindles synchronized[1], so if the host tells both drives to move to the same track and they both arrive at the same time, you still must wait for both of them to get the data, which won't happen until the appropriate region on each one has passed by the head. In practice, I think the spindle thing is a relatively small problem. Actually, both of them are probably relatively small, but the point is that striping gains you something when you are doing big bulk transfers, but it's not totally without cost. - Logan [1] It's been a while, but I believe server-class SCSI drives used to have a connector you could use to synchronize the spindles of several drives. |
#26
|
|||
|
|||
"Logan Shaw" wrote in message
Arny Krueger wrote: "Logan Shaw" wrote in message By the way, I have never seen a system that actually does striping the way you describe, even though it is a useful idea. Interesting. I've never seen one that doesn't do it this way. Vendors don't seem to talk much about this, but it explains why striping a pair of disks can yield a stripe set that runs slower than either disck, when the disks are mostly empty. There are a couple of miscellaneous reasons that a striped set can run slower than an individual disk. The first reason is that both disk heads may have to seek to satisfy an I/O. Let's say you want to read a small amount of data, and it stretches across the boundary from one stripe to the next. If you had just one drive, it would seek to the beginning of your data and 99% chance it would read it all in one continuous transfer. I'm including sequential reads on a defregged disks. But with striped disks, both disks' heads are in the wrong position (probably), and both have to move to the right position and read before you get back the data you want. Aagin, not a prolblem with sequential reads on a defragged disc. |
#27
|
|||
|
|||
"Rich.Andrews" wrote in message
I don't know where Arny got the idea that disks are striped as described, but that isn't how it is done in the real world. Sez you. On a write it would not matter one bit if one disc was writing from the outside in or inside out. Not true. Writes on the inside tracks are far slower than writes on the outside track. I can't imagine how anybody can call this mattering not one bit! The point of striping is to improve speed and improve speed it does dramatically. Of course. However, the speed of a disk also varies dramatically as it fills up and data is written closer and closer to the inside of the disk. I have seen large arrays write a gigabyte of data per second and there was none of that business of one disk starting at block 0 and the other starting at All fine and good, but in the context of a discussion of how striping can even out the performance of a disk, its a red herring. |
#28
|
|||
|
|||
"Arny Krueger" wrote in
: "Rich.Andrews" wrote in message I don't know where Arny got the idea that disks are striped as described, but that isn't how it is done in the real world. Sez you. On a write it would not matter one bit if one disc was writing from the outside in or inside out. Not true. Writes on the inside tracks are far slower than writes on the outside track. I can't imagine how anybody can call this mattering not one bit! The point of striping is to improve speed and improve speed it does dramatically. Of course. However, the speed of a disk also varies dramatically as it fills up and data is written closer and closer to the inside of the disk. I have seen large arrays write a gigabyte of data per second and there was none of that business of one disk starting at block 0 and the other starting at All fine and good, but in the context of a discussion of how striping can even out the performance of a disk, its a red herring. Arny, There is something called interleaving that is done to take care of any potential geometry related speed issues. I worked for AT&T Computer Systems in the SCSI lab for a while. I know how many raid controller work including hardware and software raid solutions and none of them that we worked with ever did that "complementary striping" you mention. Even current software raid solutions like Veritas and Sun do not do that. That also includes hardware solutions like EMC, DPT, Adaptec, and a few others. We never had any issues with data transfer rates no matter if it was at the beginning or the end of the discs. When I mean issues, I mean the speed did not change end to end. The limiting factor is all cases was the speed of the SCSI bus, not the disks and that includes fiber attached controllers. Now if you can find a disc that does not change its interleaving from end to end, I suggest you throw that POS away and buy something else. Might I suggest you do some additional research with a review of some pertinent source code. Review some hard drive conroller software if you can find it. The interleaving maps are there. Sorry Arny, in this case you are clearly wrong. r -- Nothing beats the bandwidth of a station wagon filled with DLT tapes. |
Reply |
Thread Tools | |
Display Modes | |
|
|
Similar Threads | ||||
Thread | Forum | |||
iPod network Aux interface? | Tech | |||
Will a network drive keep up? | Pro Audio | |||
802.11b/g network interface for stereo? | Tech | |||
Computer guys help needed- piping audio thru network.... | Pro Audio | |||
The Car Audio Network (CAN) | Car Audio |