Re:SerialATA doesn't seem very advanced
on
Serial ATA Coming
·
· Score: 1
Ah, it isn't the distance that is important, but the ability of the signalling layer to have external quality connectors, and EMI tolerances. While differential, you still have to worry about EMI creation. Further, 1m is the total distance, without repeaters and external termination, that is 1m from the chip to the disks chip, or close to it. Most SCSI connectors are run almost a foot across a card/mb, there goes one foot. Then you account for the fact that there is probably another 6" lost going to the device via connectors, which btw more than 2 of lower your distance to around 2 feet. Once you add up all the extra impediance changes and the board run lenghts, you really have between 1 foot an 18 inchs of external connect, assuming a cable which is rated for out of box connections and some form of connector specificaion.
Given that the physical layer of sATA is very similar to that of FC, there is no reason that the signals could not be boosted to do 15ms though. The big question is what would you hook up external, with out a hub of some kind, which the sATA 1.0 spec most definitely does not address, there is only the possability for one external device.
Re:SerialATA doesn't seem very advanced
on
Serial ATA Coming
·
· Score: 2, Insightful
From 1, you can't hot swap with a cable. Hot-swap requires a backplane. How you get power to the backplane is up to you. They have two connector styles. One standardizes the location of the old style power connectors, then they provide a single connector for power/signal, just like SCSI did, but they do not run power off of the adapter card and into the signal cable. If you are designing a backplane for drives, then you'd use the same connector, but that is different from the same cable.
2. This is incorrect, the signalling is shorter, and sATA does not want to compete with SCSI, in the future, SCSI and sATA will use the same interconnect and you choose based on how much money you want to spend. SCSI runs 25 meters point to point while sATA is 1m.
3. Nope, the distance is still 1m. No extra length. It is possible to change the spec a bit and give distance, but what would you run it to. Since it is point-to-point, and the software stacks can't handle more than two devices on a bus, what would you connect but one drive, or possibly two with a special chip?
SCSI is offering a compatible solution at http://www.serialattachedscsi.com
Although the signals look a lot like FC, it is only intended to be used internally, over short distances.
It also has all the same software layers as IDE, and is mainly aimed at the low-end and commodity markets.
The main benefits being hot-swap and the much smaller cables (only 2 conductors as opposed to 40/80). This will help cabling inside of cases and air-flow.
Since the group which developed the E10k has been purchased by Sun, they have been working on their next generation of the *fire systems, which will be remarkably faster than the 10k. That combined with the new Ultra 3 and 4 processors will probably take the title back within about 6 months
A number of things make SSD superior to disk caching.
1. Latency -- A cache is just that, cached. It must be loaded initially, and this takes time. A SSD has orders of magnitude lower latency and can load a cache much faster. Disks caches are only good if you are reusing data. Applications like databases and stream applications don't take advantage of caching.
2. Bandwidth -- While you can get bandwidth equal to that or SSDs out of Raid, memory has so much more bandwidth left over it can be used across several machines. Checkout http://www.texmemsys.com they have a SSD with upto 3.2GB/s of I/O, that more than the I/O subsystems of most systems.
3. Size -- Many computers have realistic limitations on I/O, disk subsystems are less contrained. Most desktop/workstations top out at 4-8GBs. You can add SSDs until your head spins.
4. Volatiliy, while SSDs are volatile, most have battery backups, or can be mounted externally to other redundant power supplies. Its much easier to recover memory viewed as a disk from an OS standpoint than it is to recover it from system memory.
In the end, the algorithm being processed is more critical to which platform is best, SMP, clustering, or NUMA.... based on it being I/O, bandwidth, or processor limitted.
In any event, if your software doesn't support multiple processes or threads, the issue is moot, and a UP box will be only nominally slower than an SMP box of the same speed. Ask most gamers if they would rather have a dual 600MHz machine, or a UP 800 MHz.
SMP does have some disadvantages over a cluster architecture if say the algorithm is extremly partionable but memory bandwidth limited. SMP boxes tend to share processor bandwidth, so processor utilization goes down after the memory is saturated. If you can send half the job off to another system, and have two UP machines run a bandwidth limited application, it will run much faster than on the same processors in one box. Now, this is a rare situation, but can still occur depending on the algorithm.
I am not sure how parallizable MIDI effects are, but earlier post indicated it was I/O limitted.
Of course, getting applications to support DSP/vector processor nodes of some sort would be the best solution.
Unfortunately this would not be true.
An military jumper has already broken the sound barrier in free fall jumping from a balloon back in the 50s or 60s.
Apple and Motorola repeatedly turned down attempts by IBM to share the burden of producing these chips. These Power4, which are really just 64-bit enhanced PowerPCs, are produced for IBMs enterprise server division. For political reasons, I seriously doubt IBM will sell these processors in anything other than a complete system, they have to date been totally unwilling to share these chips with anyone who wouldn't kick the cash in to sell them.
On a side note, I have worked for IBM in the past, and know those who work for them still, and this chip has gotten an enourmous amount of resources. To the point it has become one of the largest projects IBM has ever executed. The processor team is rather large, so I doubt anyone at IBM would consider this chip Ho-Hum by any measure.
While the industry and others are moving towards SMT architectures rapidly, for once in recent history, IBM seems to hold the current lead.
One would wonder though, why you looked to Compaq for the HBAs when there are plenty of other HBAs out there. It should have had no difference in the connection to the Storage Works SAN system.
One of the problems with 21x64 and thus Athlon chipsets is that the busses are designed to handle high front side bus speeds. To do this, the bus is a point to point bus with the processor connecting directly to the chipset. This translates to more wires and pins on SMP chipsets. The issue is further complicated by caching issues. Alphas use an L3 cache which the chipset controls to share data amoungst SMP nodes. This increases the compexity of SMP chipsets for the architecture. Other people have also mentions, and correctly from what I have heard, that AMD had no desire to produce chipsets, and this has probably been a reason they where slower to ramp up additional higher end chipsets. VIA might very well do similar sets, but I imagine AMD does not want to place all its eggs in that basket.
To some degree this is correct, but for many applications this is not.
As some people have mentioned, applications which are write intensive will be bandwidth limited by the syncronizations to the physical media, creation of temp files and the like.
Another problem with disk cache is that it is much less deterministic for real time considerations.
Systems with heavy load also have a tendancy to thrash disk cache and migrating critical files to a disk cache might be a better idea.
Of course, there is always the option of solid state disks, which provide many of the benefits of ramdisk and few of the disadvantages in hard disks.
Slashdot Mirror
Ah, it isn't the distance that is important, but the ability of the signalling layer to have external quality connectors, and EMI tolerances. While differential, you still have to worry about EMI creation. Further, 1m is the total distance, without repeaters and external termination, that is 1m from the chip to the disks chip, or close to it. Most SCSI connectors are run almost a foot across a card/mb, there goes one foot. Then you account for the fact that there is probably another 6" lost going to the device via connectors, which btw more than 2 of lower your distance to around 2 feet. Once you add up all the extra impediance changes and the board run lenghts, you really have between 1 foot an 18 inchs of external connect, assuming a cable which is rated for out of box connections and some form of connector specificaion.
Given that the physical layer of sATA is very similar to that of FC, there is no reason that the signals could not be boosted to do 15ms though. The big question is what would you hook up external, with out a hub of some kind, which the sATA 1.0 spec most definitely does not address, there is only the possability for one external device.
From 1, you can't hot swap with a cable. Hot-swap requires a backplane. How you get power to the backplane is up to you. They have two connector styles. One standardizes the location of the old style power connectors, then they provide a single connector for power/signal, just like SCSI did, but they do not run power off of the adapter card and into the signal cable. If you are designing a backplane for drives, then you'd use the same connector, but that is different from the same cable.
2. This is incorrect, the signalling is shorter, and sATA does not want to compete with SCSI, in the future, SCSI and sATA will use the same interconnect and you choose based on how much money you want to spend. SCSI runs 25 meters point to point while sATA is 1m.
3. Nope, the distance is still 1m. No extra length. It is possible to change the spec a bit and give distance, but what would you run it to. Since it is point-to-point, and the software stacks can't handle more than two devices on a bus, what would you connect but one drive, or possibly two with a special chip?
SCSI is offering a compatible solution at http://www.serialattachedscsi.com
Currently there is a 1m cable length.
Although the signals look a lot like FC, it is only intended to be used internally, over short distances.
It also has all the same software layers as IDE, and is mainly aimed at the low-end and commodity markets.
The main benefits being hot-swap and the much smaller cables (only 2 conductors as opposed to 40/80). This will help cabling inside of cases and air-flow.
It is both 128GB and 137GBs. Depends on if you believe a drive vendor and a GB is 10^9 bytes (137.439...GB), or you believe it is 128*1024^3, (128GB).
;)
Still smaller than SCSI drives
Just a small note, IBM does not actually produce their own external disk arrays other than the Shark line. Their systems are silk screened.
Since the group which developed the E10k has been purchased by Sun, they have been working on their next generation of the *fire systems, which will be remarkably faster than the 10k. That combined with the new Ultra 3 and 4 processors will probably take the title back within about 6 months
Considering they have been in it for at least 20 years, I'd say awhile. They are migrating away from the S/390 machines to Sun compatible machines.
A number of things make SSD superior to disk caching. 1. Latency -- A cache is just that, cached. It must be loaded initially, and this takes time. A SSD has orders of magnitude lower latency and can load a cache much faster. Disks caches are only good if you are reusing data. Applications like databases and stream applications don't take advantage of caching. 2. Bandwidth -- While you can get bandwidth equal to that or SSDs out of Raid, memory has so much more bandwidth left over it can be used across several machines. Checkout http://www.texmemsys.com they have a SSD with upto 3.2GB/s of I/O, that more than the I/O subsystems of most systems. 3. Size -- Many computers have realistic limitations on I/O, disk subsystems are less contrained. Most desktop/workstations top out at 4-8GBs. You can add SSDs until your head spins. 4. Volatiliy, while SSDs are volatile, most have battery backups, or can be mounted externally to other redundant power supplies. Its much easier to recover memory viewed as a disk from an OS standpoint than it is to recover it from system memory.
In the end, the algorithm being processed is more critical to which platform is best, SMP, clustering, or NUMA.... based on it being I/O, bandwidth, or processor limitted. In any event, if your software doesn't support multiple processes or threads, the issue is moot, and a UP box will be only nominally slower than an SMP box of the same speed. Ask most gamers if they would rather have a dual 600MHz machine, or a UP 800 MHz. SMP does have some disadvantages over a cluster architecture if say the algorithm is extremly partionable but memory bandwidth limited. SMP boxes tend to share processor bandwidth, so processor utilization goes down after the memory is saturated. If you can send half the job off to another system, and have two UP machines run a bandwidth limited application, it will run much faster than on the same processors in one box. Now, this is a rare situation, but can still occur depending on the algorithm. I am not sure how parallizable MIDI effects are, but earlier post indicated it was I/O limitted. Of course, getting applications to support DSP/vector processor nodes of some sort would be the best solution.
Oops, correct, speed read
Unfortunately this would not be true. An military jumper has already broken the sound barrier in free fall jumping from a balloon back in the 50s or 60s.
Apple and Motorola repeatedly turned down attempts by IBM to share the burden of producing these chips. These Power4, which are really just 64-bit enhanced PowerPCs, are produced for IBMs enterprise server division. For political reasons, I seriously doubt IBM will sell these processors in anything other than a complete system, they have to date been totally unwilling to share these chips with anyone who wouldn't kick the cash in to sell them. On a side note, I have worked for IBM in the past, and know those who work for them still, and this chip has gotten an enourmous amount of resources. To the point it has become one of the largest projects IBM has ever executed. The processor team is rather large, so I doubt anyone at IBM would consider this chip Ho-Hum by any measure. While the industry and others are moving towards SMT architectures rapidly, for once in recent history, IBM seems to hold the current lead.
One would wonder though, why you looked to Compaq for the HBAs when there are plenty of other HBAs out there. It should have had no difference in the connection to the Storage Works SAN system.
One of the problems with 21x64 and thus Athlon chipsets is that the busses are designed to handle high front side bus speeds. To do this, the bus is a point to point bus with the processor connecting directly to the chipset. This translates to more wires and pins on SMP chipsets. The issue is further complicated by caching issues. Alphas use an L3 cache which the chipset controls to share data amoungst SMP nodes. This increases the compexity of SMP chipsets for the architecture. Other people have also mentions, and correctly from what I have heard, that AMD had no desire to produce chipsets, and this has probably been a reason they where slower to ramp up additional higher end chipsets. VIA might very well do similar sets, but I imagine AMD does not want to place all its eggs in that basket.
To some degree this is correct, but for many applications this is not. As some people have mentioned, applications which are write intensive will be bandwidth limited by the syncronizations to the physical media, creation of temp files and the like. Another problem with disk cache is that it is much less deterministic for real time considerations. Systems with heavy load also have a tendancy to thrash disk cache and migrating critical files to a disk cache might be a better idea. Of course, there is always the option of solid state disks, which provide many of the benefits of ramdisk and few of the disadvantages in hard disks.