Re: [PERFORM] [PATCHES] [HACKERS] ARC Memory Usage analysis
On Wed, 2004-10-27 at 01:39, Josh Berkus wrote: Thomas, As a result, I was intending to inflate the value of effective_cache_size to closer to the amount of unused RAM on some of the machines I admin (once I've verified that they all have a unified buffer cache). Is that correct? Currently, yes. I now believe the answer to that is no, that is not fully correct, following investigation into how to set that parameter correctly. Right now, e_c_s is used just to inform the planner and make index vs. table scan and join order decisions. Yes, I agree that is what e_c_s is used for. ...lets go deeper: effective_cache_size is used to calculate the number of I/Os required to index scan a table, which varies according to the size of the available cache (whether this be OS cache or shared_buffers). The reason to do this is because whether a table is in cache can make a very great difference to access times; *small* tables tend to be the ones that vary most significantly. PostgreSQL currently uses the Mackert and Lohman [1989] equation to assess how much of a table is in cache in a blocked DBMS with a finite cache. The Mackert and Lohman equation is accurate, as long as the parameter b is reasonably accurately set. [I'm discussing only the current behaviour here, not what it can or should or could be] If it is incorrectly set, then the equation will give the wrong answer for small tables. The same answer (i.e. same asymptotic behaviour) is returned for very large tables, but they are the ones we didn't worry about anyway. Getting the equation wrong means you will choose sub-optimal plans, potentially reducing your performance considerably. As I read it, effective_cache_size is equivalent to the parameter b, defined as (p.3) minimum buffer size dedicated to a given scan. ML they point out (p.3) We...do not consider interactions of multiple users sharing the buffer for multiple file accesses. Either way, ML aren't talking about the total size of the cache, which we would interpret to mean shared_buffers + OS cache, in our effort to not forget the beneficial effect of the OS cache. They use the phrase dedicated to a given scan AFAICS effective_cache_size should be set to a value that reflects how many other users of the cache there might be. If you know for certain you're the only user, set it according to the existing advice. If you know you aren't, then set it an appropriate factor lower. Setting that accurately on a system wide basis may clearly be difficult and setting it high will often be inappropriate. The manual is not clear as to how to set effective_cache_size. Other advice misses out the effect of the many scans/many tables issue and will give the wrong answer for many calculations, and thus produce incorrect plans for 8.0 (and earlier releases also). This is something that needs to be documented rather than a bug fix. It's a complex one, so I'll await all of your objections before I write a new doc patch. [Anyway, I do hope I've missed something somewhere in all that, though I've read their paper twice now. Fairly accessible, but requires interpretation to the PostgreSQL case. Mackert and Lohman [1989] Index Scans using a finite LRU buffer: A validated I/O model] The problem which Simon is bringing up is part of a discussion about doing *more* with the information supplied by e_c_s.He points out that it's not really related to the *real* probability of any particular table being cached. At least, if I'm reading him right. Yes, that was how Jan originally meant to discuss it, but not what I meant. Best regards, Simon Riggs ---(end of broadcast)--- TIP 8: explain analyze is your friend
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
On 10/26/2004 1:53 AM, Tom Lane wrote: Greg Stark [EMAIL PROTECTED] writes: Tom Lane [EMAIL PROTECTED] writes: Another issue is what we do with the effective_cache_size value once we have a number we trust. We can't readily change the size of the ARC lists on the fly. Huh? I thought effective_cache_size was just used as an factor the cost estimation equation. Today, that is true. Jan is speculating about using it as a parameter of the ARC cache management algorithm ... and that worries me. If we need another config option, it's not that we are running out of possible names, is it? Jan -- #==# # It's easier to get forgiveness for being wrong than for being right. # # Let's break this rule - forgive me. # #== [EMAIL PROTECTED] # ---(end of broadcast)--- TIP 1: subscribe and unsubscribe commands go to [EMAIL PROTECTED]
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
Jan Wieck [EMAIL PROTECTED] writes: On 10/26/2004 1:53 AM, Tom Lane wrote: Greg Stark [EMAIL PROTECTED] writes: Tom Lane [EMAIL PROTECTED] writes: Another issue is what we do with the effective_cache_size value once we have a number we trust. We can't readily change the size of the ARC lists on the fly. Huh? I thought effective_cache_size was just used as an factor the cost estimation equation. Today, that is true. Jan is speculating about using it as a parameter of the ARC cache management algorithm ... and that worries me. If we need another config option, it's not that we are running out of possible names, is it? No, the point is that the value is not very trustworthy at the moment. regards, tom lane ---(end of broadcast)--- TIP 4: Don't 'kill -9' the postmaster
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
On Mon, Oct 25, 2004 at 05:53:25PM -0400, Tom Lane wrote: Greg Stark [EMAIL PROTECTED] writes: So I would suggest using something like 100us as the threshold for determining whether a buffer fetch came from cache. I see no reason to hardwire such a number. On any hardware, the distribution is going to be double-humped, and it will be pretty easy to determine a cutoff after minimal accumulation of data. The real question is whether we can afford a pair of gettimeofday() calls per read(). This isn't a big issue if the read actually results in I/O, but if it doesn't, the percentage overhead could be significant. How invasive would reading the CPU counter be, if it is available? A read operation should avoid flushing a cache line and we can throw out the obvious outliers since we only need an estimate and not the actual value. --Ken ---(end of broadcast)--- TIP 1: subscribe and unsubscribe commands go to [EMAIL PROTECTED]
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
Tom Lane wrote: Greg Stark [EMAIL PROTECTED] writes: So I would suggest using something like 100us as the threshold for determining whether a buffer fetch came from cache. I see no reason to hardwire such a number. On any hardware, the distribution is going to be double-humped, and it will be pretty easy to determine a cutoff after minimal accumulation of data. The real question is whether we can afford a pair of gettimeofday() calls per read(). This isn't a big issue if the read actually results in I/O, but if it doesn't, the percentage overhead could be significant. If we assume that the effective_cache_size value isn't changing very fast, maybe it would be good enough to instrument only every N'th read (I'm imagining N on the order of 100) for this purpose. Or maybe we need only instrument reads that are of blocks that are close to where the ARC algorithm thinks the cache edge is. If it's decided to instrument reads, then perhaps an even better use of it would be to tune random_page_cost. If the storage manager knows the difference between a sequential scan and a random scan, then it should easily be able to measure the actual performance it gets for each and calculate random_page_cost based on the results. While the ARC lists can't be tuned on the fly, random_page_cost can. One small problem is that the time measurement gives you only a lower bound on the time the read() actually took. In a heavily loaded system you might not get the CPU back for long enough to fool you about whether the block came from cache or not. True, but that's information that you'd want to factor into the performance measurements anyway. The database needs to know how much wall clock time it takes for it to fetch a page under various circumstances from disk via the OS. For determining whether or not the read() hit the disk instead of just OS cache, what would matter is the average difference between the two. That's admittedly a problem if the difference is less than the noise, though, but at the same time that would imply that given the circumstances it really doesn't matter whether or not the page was fetched from disk: the difference is small enough that you could consider them equivalent. You don't need 100% accuracy for this stuff, just statistically significant accuracy. Another issue is what we do with the effective_cache_size value once we have a number we trust. We can't readily change the size of the ARC lists on the fly. Compare it with the current value, and notify the DBA if the values are significantly different? Perhaps write the computed value to a file so the DBA can look at it later? Same with other values that are computed on the fly. In fact, it might make sense to store them in a table that gets periodically updated, and load their values from that table, and then the values in postgresql.conf or the command line would be the default that's used if there's nothing in the table (and if you really want fine-grained control of this process, you could stick a boolean column in the table to indicate whether or not to load the value from the table at startup time). -- Kevin Brown [EMAIL PROTECTED] ---(end of broadcast)--- TIP 1: subscribe and unsubscribe commands go to [EMAIL PROTECTED]
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
On Tue, 26 Oct 2004, Greg Stark wrote: I see mmap or O_DIRECT being the only viable long-term stable states. My natural inclination was the former but after the latest thread on the subject I suspect it'll be forever out of reach. That makes O_DIRECT And a Postgres managed cache the only real choice. Having both caches is just a waste of memory and a waste of cpu cycles. I don't see why mmap is any more out of reach than O_DIRECT; it's not all that much harder to implement, and mmap (and madvise!) is more widely available. But if using two caches is only costing us 1% in performance, there's not really much point cjs -- Curt Sampson [EMAIL PROTECTED] +81 90 7737 2974 http://www.NetBSD.org Make up enjoying your city life...produced by BIC CAMERA ---(end of broadcast)--- TIP 2: you can get off all lists at once with the unregister command (send unregister YourEmailAddressHere to [EMAIL PROTECTED])
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
On Mon, 2004-10-25 at 16:34, Jan Wieck wrote: The problem is, with a too small directory ARC cannot guesstimate what might be in the kernel buffers. Nor can it guesstimate what recently was in the kernel buffers and got pushed out from there. That results in a way too small B1 list, and therefore we don't get B1 hits when in fact the data was found in memory. B1 hits is what increases the T1target, and since we are missing them with a too small directory size, our implementation of ARC is propably using a T2 size larger than the working set. That is not optimal. I think I have seen that the T1 list shrinks too much, but need more tests...with some good test results The effectiveness of ARC relies upon the balance between the often conflicting requirements of recency and frequency. It seems possible, even likely, that pgsql's version of ARC may need some subtle changes to rebalance it - if we are unlikely enough to find cases where it genuinely is out of balance. Many performance tests are required, together with a few ideas on extra parameters to includehence my support of Jan's ideas. That's also why I called the B1+B2 hit ratio turbulence because it relates to how much oscillation is happening between T1 and T2. In physical systems, we expect the oscillations to be damped, but there is no guarantee that we have a nearly critically damped oscillator. (Note that the absence of turbulence doesn't imply that T1+T2 is optimally sized, just that is balanced). [...and all though the discussion has wandered away from my original patch...would anybody like to commit, or decline the patch?] If we would replace the dynamic T1 buffers with a max_backends*2 area of shared buffers, use a C value representing the effective cache size and limit the T1target on the lower bound to effective cache size - shared buffers, then we basically moved the T1 cache into the OS buffers. Limiting the minimum size of T1len to be 2* maxbackends sounds like an easy way to prevent overbalancing of T2, but I would like to follow up on ways to have T1 naturally stay larger. I'll do a patch with this idea in, for testing. I'll call this T1 minimum size so we can discuss it. Any other patches are welcome... It could be that B1 is too small and so we could use a larger value of C to keep track of more blocks. I think what is being suggested is two GUCs: shared_buffers (as is), plus another one, larger, which would allow us to track what is in shared_buffers and what is in OS cache. I have comments on effective cache size below On Mon, 2004-10-25 at 17:03, Tom Lane wrote: Jan Wieck [EMAIL PROTECTED] writes: This all only holds water, if the OS is allowed to swap out shared memory. And that was my initial question, how likely is it to find this to be true these days? I think it's more likely that not that the OS will consider shared memory to be potentially swappable. On some platforms there is a shmctl call you can make to lock your shmem in memory, but (a) we don't use it and (b) it may well require privileges we haven't got anyway. Are you saying we shouldn't, or we don't yet? I simply assumed that we did use that function - surely it must be at least an option? RHEL supports this at least It may well be that we don't have those privileges, in which case we turn off the option. Often, we (or I?) will want to install a dedicated server, so we should have all the permissions we need, in which case... This has always been one of the arguments against making shared_buffers really large, of course --- if the buffers aren't all heavily used, and the OS decides to swap them to disk, you are worse off than you would have been with a smaller shared_buffers setting. Not really, just an argument against making them *too* large. Large *and* utilised is OK, so we need ways of judging optimal sizing. However, I'm still really nervous about the idea of using effective_cache_size to control the ARC algorithm. That number is usually entirely bogus. Right now it is only a second-order influence on certain planner estimates, and I am afraid to rely on it any more heavily than that. ...ah yes, effective_cache_size. The manual describes effective_cache_size as if it had something to do with the OS, and some of this discussion has picked up on that. effective_cache_size is used in only two places in the code (both in the planner), as an estimate for calculating the cost of a) nonsequential access and b) index access, mainly as a way of avoiding overestimates of access costs for small tables. There is absolutely no implication in the code that effective_cache_size measures anything in the OS; what it gives is an estimate of the number of blocks that will be available from *somewhere* in memory (i.e. in shared_buffers OR OS cache) for one particular table (the one currently being considered by the planner). Crucially, the size referred to is the size of the *estimate*, not the size
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
On Tue, 2004-10-26 at 06:53, Tom Lane wrote: Greg Stark [EMAIL PROTECTED] writes: Tom Lane [EMAIL PROTECTED] writes: Another issue is what we do with the effective_cache_size value once we have a number we trust. We can't readily change the size of the ARC lists on the fly. Huh? I thought effective_cache_size was just used as an factor the cost estimation equation. Today, that is true. Jan is speculating about using it as a parameter of the ARC cache management algorithm ... and that worries me. ISTM that we should be optimizing the use of shared_buffers, not whats outside. Didn't you (Tom) already say that? BTW, very good ideas on how to proceed, but why bother? For me, if the sysadmin didn't give shared_buffers to PostgreSQL, its because the memory is intended for use by something else and so not available at all. At least not dependably. The argument against large shared_buffers because of swapping applies to that assumption also...the OS cache is too volatile to attempt to gauge sensibly. There's an argument for improving performance for people that haven't set their parameters correctly, but thats got to be a secondary consideration anyhow. -- Best Regards, Simon Riggs ---(end of broadcast)--- TIP 1: subscribe and unsubscribe commands go to [EMAIL PROTECTED]
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
On Tue, 2004-10-26 at 09:49, Simon Riggs wrote: On Mon, 2004-10-25 at 16:34, Jan Wieck wrote: The problem is, with a too small directory ARC cannot guesstimate what might be in the kernel buffers. Nor can it guesstimate what recently was in the kernel buffers and got pushed out from there. That results in a way too small B1 list, and therefore we don't get B1 hits when in fact the data was found in memory. B1 hits is what increases the T1target, and since we are missing them with a too small directory size, our implementation of ARC is propably using a T2 size larger than the working set. That is not optimal. I think I have seen that the T1 list shrinks too much, but need more tests...with some good test results If we would replace the dynamic T1 buffers with a max_backends*2 area of shared buffers, use a C value representing the effective cache size and limit the T1target on the lower bound to effective cache size - shared buffers, then we basically moved the T1 cache into the OS buffers. Limiting the minimum size of T1len to be 2* maxbackends sounds like an easy way to prevent overbalancing of T2, but I would like to follow up on ways to have T1 naturally stay larger. I'll do a patch with this idea in, for testing. I'll call this T1 minimum size so we can discuss it. Don't know whether you've seen this latest update on the ARC idea: Sorav Bansal and Dharmendra S. Modha, CAR: Clock with Adaptive Replacement, in Proceedings of the USENIX Conference on File and Storage Technologies (FAST), pages 187--200, March 2004. [I picked up the .pdf here http://citeseer.ist.psu.edu/bansal04car.html] In that paper Bansal and Modha introduce an update to ARC called CART which they say is more appropriate for databases. Their idea is to introduce a temporal locality window as a way of making sure that blocks called twice within a short period don't fall out of T1, though don't make it into T2 either. Strangely enough the temporal locality window is made by increasing the size of T1... in an adpative way, of course. If we were going to put a limit on the minimum size of T1, then this would put a minimal temporal locality window in placerather than the increased complexity they go to in order to make T1 larger. I note test results from both the ARC and CAR papers that show that T2 usually represents most of C, so the observations that T1 is very small is not atypical. That implies that the cost of managing the temporal locality window in CART is usually wasted, even though it does cut in as an overall benefit: The results show that CART is better than ARC over the whole range of cache sizes tested (16MB to 4GB) and workloads (apart from 1 out 22). If we were to implement a minimum size of T1, related as suggested to number of users, then this would provide a reasonable approximation of the temporal locality window. This wouldn't prevent the adaptation of T1 to be higher than this when required. Jan has already optimised ARC for PostgreSQL by the addition of a special lookup on transactionId required to optimise for the double cache lookup of select/update that occurs on a T1 hit. That seems likely to be able to be removed as a result of having a larger T1. I'd suggest limiting T1 to be a value of: shared_buffers = 1000 T1limit = max_backends *0.75 shared_buffers = 2000 T1limit = max_backends shared_buffers = 5000 T1limit = max_backends *1.5 shared_buffers 5000 T1limit = max_backends *2 I'll try some tests with both - minimum size of T1 - update optimisation removed Thoughts? -- Best Regards, Simon Riggs ---(end of broadcast)--- TIP 1: subscribe and unsubscribe commands go to [EMAIL PROTECTED]
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
Curt Sampson [EMAIL PROTECTED] writes: On Tue, 26 Oct 2004, Greg Stark wrote: I see mmap or O_DIRECT being the only viable long-term stable states. My natural inclination was the former but after the latest thread on the subject I suspect it'll be forever out of reach. That makes O_DIRECT And a Postgres managed cache the only real choice. Having both caches is just a waste of memory and a waste of cpu cycles. I don't see why mmap is any more out of reach than O_DIRECT; it's not all that much harder to implement, and mmap (and madvise!) is more widely available. Because there's no way to prevent a write-out from occurring and no way to be notified by mmap before a write-out occurs, and Postgres wants to do its WAL logging at that time if it hasn't already happened. But if using two caches is only costing us 1% in performance, there's not really much point Well firstly it depends on the work profile. It can probably get much higher than we saw in that profile if your work load is causing more fresh buffers to be fetched. Secondly it also reduces the amount of cache available. If you have 256M of ram with about 200M free, and 40Mb of ram set aside for Postgres's buffer cache then you really only get 160Mb. It's costing you 20% of your cache, and reducing the cache hit rate accordingly. Thirdly the kernel doesn't know as much as Postgres about the load. Postgres could optimize its use of cache based on whether it knows the data is being loaded by a vacuum or sequential scan rather than an index lookup. In practice Postgres has gone with ARC which I suppose a kernel could implement anyways, but afaik neither linux nor BSD choose to do anything like it. -- greg ---(end of broadcast)--- TIP 9: the planner will ignore your desire to choose an index scan if your joining column's datatypes do not match
Re: [PERFORM] [PATCHES] [HACKERS] ARC Memory Usage analysis
Thomas, As a result, I was intending to inflate the value of effective_cache_size to closer to the amount of unused RAM on some of the machines I admin (once I've verified that they all have a unified buffer cache). Is that correct? Currently, yes. Right now, e_c_s is used just to inform the planner and make index vs. table scan and join order decisions. The problem which Simon is bringing up is part of a discussion about doing *more* with the information supplied by e_c_s.He points out that it's not really related to the *real* probability of any particular table being cached. At least, if I'm reading him right. -- --Josh Josh Berkus Aglio Database Solutions San Francisco ---(end of broadcast)--- TIP 4: Don't 'kill -9' the postmaster
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
On Wed, 26 Oct 2004, Greg Stark wrote: I don't see why mmap is any more out of reach than O_DIRECT; it's not all that much harder to implement, and mmap (and madvise!) is more widely available. Because there's no way to prevent a write-out from occurring and no way to be notified by mmap before a write-out occurs, and Postgres wants to do its WAL logging at that time if it hasn't already happened. I already described a solution to that problem in a post earlier in this thread (a write queue on the block). I may even have described it on this list a couple of years ago, that being about the time I thought it up. (The mmap idea just won't die, but at least I wasn't the one to bring it up this time. :-)) Well firstly it depends on the work profile. It can probably get much higher than we saw in that profile True, but 1% was is much, much lower than I'd expected. That tells me that my intuitive idea of the performance model is wrong, which means, for me at least, it's time to shut up or put up some benchmarks. Secondly it also reduces the amount of cache available. If you have 256M of ram with about 200M free, and 40Mb of ram set aside for Postgres's buffer cache then you really only get 160Mb. It's costing you 20% of your cache, and reducing the cache hit rate accordingly. Yeah, no question about that. Thirdly the kernel doesn't know as much as Postgres about the load. Postgres could optimize its use of cache based on whether it knows the data is being loaded by a vacuum or sequential scan rather than an index lookup. In practice Postgres has gone with ARC which I suppose a kernel could implement anyways, but afaik neither linux nor BSD choose to do anything like it. madvise(). cjs -- Curt Sampson [EMAIL PROTECTED] +81 90 7737 2974 http://www.NetBSD.org Make up enjoying your city life...produced by BIC CAMERA ---(end of broadcast)--- TIP 8: explain analyze is your friend
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
On Mon, 2004-10-25 at 23:53, Tom Lane wrote: Greg Stark [EMAIL PROTECTED] writes: Tom Lane [EMAIL PROTECTED] writes: Another issue is what we do with the effective_cache_size value once we have a number we trust. We can't readily change the size of the ARC lists on the fly. Huh? I thought effective_cache_size was just used as an factor the cost estimation equation. Today, that is true. Jan is speculating about using it as a parameter of the ARC cache management algorithm ... and that worries me. Because it's so often set wrong I take it. But if it's set right, and it makes the the database faster to pay attention to it, then I'd be in favor of it. Or at least having a switch to turn on the ARC buffer's ability to look at it. Or is it some other issue, having to do with the idea of knowing effective cache size cause a positive effect overall on the ARC algorhythm? ---(end of broadcast)--- TIP 6: Have you searched our list archives? http://archives.postgresql.org
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
Jan Wieck [EMAIL PROTECTED] writes: This all only holds water, if the OS is allowed to swap out shared memory. And that was my initial question, how likely is it to find this to be true these days? I think it's more likely that not that the OS will consider shared memory to be potentially swappable. On some platforms there is a shmctl call you can make to lock your shmem in memory, but (a) we don't use it and (b) it may well require privileges we haven't got anyway. This has always been one of the arguments against making shared_buffers really large, of course --- if the buffers aren't all heavily used, and the OS decides to swap them to disk, you are worse off than you would have been with a smaller shared_buffers setting. However, I'm still really nervous about the idea of using effective_cache_size to control the ARC algorithm. That number is usually entirely bogus. Right now it is only a second-order influence on certain planner estimates, and I am afraid to rely on it any more heavily than that. regards, tom lane ---(end of broadcast)--- TIP 5: Have you checked our extensive FAQ? http://www.postgresql.org/docs/faqs/FAQ.html
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
Greg Stark [EMAIL PROTECTED] writes: However I wonder about another approach entirely. If postgres timed how long reads took it shouldn't find it very hard to distinguish between a cached buffer being copied and an actual i/o operation. It should be able to track the percentage of time that buffers requested are in the kernel's cache and use that directly instead of the estimated cache size. I tested this with a program that times seeking to random locations in a file. It's pretty easy to spot the break point. There are very few fetches that take between 50us and 1700us, probably they come from the drive's onboard cache. The 1700us bound probably would be lower for high end server equipment with 10k RPM drives and RAID arrays. But I doubt it will ever come close to the 100us edge, not without features like cache ram that Postgres would be better off considering to be part of effective_cache anyways. So I would suggest using something like 100us as the threshold for determining whether a buffer fetch came from cache. Here are two graphs, one showing a nice curve showing how disk seek times are distributed. It's neat to look at for that alone: inline: plot1.png This is the 1000 fastest data points zoomed to the range under 1800us: inline: plot2.png This is the program I wrote to test this: #include unistd.h #include sys/types.h #include sys/stat.h #include sys/time.h #include fcntl.h #include time.h #include stdlib.h #include stdio.h int main(int argc, char *argv[]) { int rep = atoi(argv[1]); int i; char *filename; int fd; struct stat statbuf; off_t filesize; unsigned blocksize; void *blockbuf; filename = argv[2]; fd = open(filename, O_RDONLY); fstat(fd, statbuf); filesize = statbuf.st_size; blocksize = statbuf.st_blksize; blockbuf = malloc(blocksize); srandom(getpid()^clock()); for (i=0;irep;i++) { struct timeval timeval1,timeval2; off_t offset = random()%filesize / blocksize * blocksize; gettimeofday(timeval1,NULL); lseek(fd, offset, SEEK_SET); read(fd, blockbuf, blocksize); gettimeofday(timeval2,NULL); /*printf(Read (%d at %ld) took %ld us\n, blocksize, offset, timeval2.tv_usec-timeval1.tv_usec + (timeval2.tv_sec-timeval1.tv_sec)*100);*/ printf(%ld\n, timeval2.tv_usec-timeval1.tv_usec + (timeval2.tv_sec-timeval1.tv_sec)*100); } return 0; } Here are the commands I used to generate the graphs: $ dd bs=1M count=1024 if=/dev/urandom of=/tmp/big $ ./a.out 1 /tmp/big /tmp/l $ gnuplot gnuplot set terminal png gnuplot set output /tmp/plot1.png gnuplot plot '/tmp/l2' with points pointtype 1 pointsize 1 gnuplot set output /tmp/plot2.png gnuplot plot [0:2000] [0:1000] '/tmp/l2' with points pointtype 1 pointsize 1 -- greg ---(end of broadcast)--- TIP 9: the planner will ignore your desire to choose an index scan if your joining column's datatypes do not match
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
Greg Stark [EMAIL PROTECTED] writes: So I would suggest using something like 100us as the threshold for determining whether a buffer fetch came from cache. I see no reason to hardwire such a number. On any hardware, the distribution is going to be double-humped, and it will be pretty easy to determine a cutoff after minimal accumulation of data. The real question is whether we can afford a pair of gettimeofday() calls per read(). This isn't a big issue if the read actually results in I/O, but if it doesn't, the percentage overhead could be significant. If we assume that the effective_cache_size value isn't changing very fast, maybe it would be good enough to instrument only every N'th read (I'm imagining N on the order of 100) for this purpose. Or maybe we need only instrument reads that are of blocks that are close to where the ARC algorithm thinks the cache edge is. One small problem is that the time measurement gives you only a lower bound on the time the read() actually took. In a heavily loaded system you might not get the CPU back for long enough to fool you about whether the block came from cache or not. Another issue is what we do with the effective_cache_size value once we have a number we trust. We can't readily change the size of the ARC lists on the fly. regards, tom lane ---(end of broadcast)--- TIP 6: Have you searched our list archives? http://archives.postgresql.org
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
Kenneth Marshall [EMAIL PROTECTED] writes: How invasive would reading the CPU counter be, if it is available? Invasive or not, this is out of the question; too unportable. regards, tom lane ---(end of broadcast)--- TIP 4: Don't 'kill -9' the postmaster
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
Is something broken with the list software? I'm receiving other emails from the list but I haven't received any of the mails in this thread. I'm only able to follow the thread based on the emails people are cc'ing to me directly. I think I've caught this behaviour in the past as well. Is it a misguided list software feature trying to avoid duplicates or something like that? It makes it really hard to follow threads in MUAs with good filtering since they're fragmented between two mailboxes. -- greg ---(end of broadcast)--- TIP 7: don't forget to increase your free space map settings
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
Greg Stark [EMAIL PROTECTED] writes: Tom Lane [EMAIL PROTECTED] writes: Another issue is what we do with the effective_cache_size value once we have a number we trust. We can't readily change the size of the ARC lists on the fly. Huh? I thought effective_cache_size was just used as an factor the cost estimation equation. Today, that is true. Jan is speculating about using it as a parameter of the ARC cache management algorithm ... and that worries me. regards, tom lane ---(end of broadcast)--- TIP 2: you can get off all lists at once with the unregister command (send unregister YourEmailAddressHere to [EMAIL PROTECTED])
Re: [PATCHES] [HACKERS] ARC Memory Usage analysis
On Fri, 2004-10-22 at 21:45, Tom Lane wrote: Jan Wieck [EMAIL PROTECTED] writes: What do you think about my other theory to make C actually 2x effective cache size and NOT to keep T1 in shared buffers but to assume T1 lives in the OS buffer cache? What will you do when initially fetching a page? It's not supposed to go directly into T2 on first use, but we're going to have some difficulty accessing a page that's not in shared buffers. I don't think you can equate the T1/T2 dichotomy to is in shared buffers or not. Yes, there are issues there. I want Jan to follow his thoughts through. This is important enough that its worth it - there's only a few even attempting this. You could maybe have a T3 list of pages that aren't in shared buffers anymore but we think are still in OS buffer cache, but what would be the point? It'd be a sufficiently bad model of reality as to be pretty much useless for stats gathering, I'd think. The OS cache is in many ways a wild horse, I agree. Jan is trying to think of ways to harness it, whereas I had mostly ignored it - but its there. Raw disk usage never allowed this opportunity. For high performance systems, we can assume that the OS cache is ours to play with - what will we do with it? We need to use it for some purposes, yet would like to ignore it for others. -- Best Regards, Simon Riggs ---(end of broadcast)--- TIP 1: subscribe and unsubscribe commands go to [EMAIL PROTECTED]