Re: [Neo4j] strange performance
Hey, Is it the exact same query? If not, it would be interesting to know how many Relations and Nodes have been traversed. Depending on graph topology this can differ a lot. cheers Martin Neumann On Tue, Nov 16, 2010 at 10:46 AM, Martin Grimmer martin.grim...@unister.dewrote: Hello, currently we evaluate Neo4j for one of our projects. Our tests showed some strange behaviour to us: ... at query 941/1000 (0,009410) with 1809 mb in 0,30 seconds - avg: 0,308696 - max: 54,533000 at query 942/1000 (0,009420) with 1809 mb in 0,069000 seconds - avg: 0,308441 - max: 54,533000 at query 943/1000 (0,009430) with 1809 mb in 0,057000 seconds - avg: 0,308174 - max: 54,533000 at query 944/1000 (0,009440) with 1809 mb in 0,038000 seconds - avg: 0,307888 - max: 54,533000 at query 945/1000 (0,009450) with 1809 mb in 0,63 seconds - avg: 0,308229 - max: 54,533000 at query 946/1000 (0,009460) with 1809 mb in* 60,997000 seconds* - avg: 0,372450 - max: 60,997000 ... fast again Our Neo4j database is about 9 GB in size with about 130M arcs and 15M nodes. A query gets two strings as input, these are keys for the lucene index service to get access to two nodes. Then we start an algorithm for these two nodes which determines nodes which are adjacent (by 2 steps BFS) to both restricted to only one specific arc type. I dont know why some queries are so much ( x100) slower. Maybe you are able to help me? best regards -- * Martin Grimmer * Developer, Semantic Web Project, IT Unister GmbH Barfußgässchen 11 | 04109 Leipzig Telefon: +49 (0)341 49288 5064 martin.grim...@unister.de mailto:%0a%20%20martin.grim...@unister.de0a%2520%2520martin.grim...@unister.de www.unister.de http://www.unister.de Vertretungsberechtigter Geschäftsführer: Thomas Wagner Amtsgericht Leipzig, HRB: 19056 ___ Neo4j mailing list User@lists.neo4j.org https://lists.neo4j.org/mailman/listinfo/user ___ Neo4j mailing list User@lists.neo4j.org https://lists.neo4j.org/mailman/listinfo/user
Re: [Neo4j] Are Relationships Singleton?
Hi, Neo4j is a multi graph - there can exist multiple relationships between two nodes even with the same type and direction. Each relationship has a unique ID so you can tell them apart. Whats your error message in detail? cheers Martin On Wed, Aug 18, 2010 at 9:54 PM, Marius Kubatz marius.kub...@udo.eduwrote: Hi! I have a very stupid question... Is it ensured that a relationship between two Nodes stays a singleton, even if another relationship of the same type and direction is added between those Nodes? Sometimes I get strange results when I delete nodes, that's why I ask. Thanks in advance and best regards, Marius ___ Neo4j mailing list User@lists.neo4j.org https://lists.neo4j.org/mailman/listinfo/user ___ Neo4j mailing list User@lists.neo4j.org https://lists.neo4j.org/mailman/listinfo/user
Re: [Neo4j] Stumped by performance issue in traversal - would take a month to run!
- Martin, I'm confused a bit about SSDs. I read up on them after I read your post. You said flash drives are best, but I read that even the highest performing flash drives are about 30MB/s read, whereas modern hard drives are at least 50MB/s. True SSDs claim to be 50MB/s too but they're quite expensive. So why is a flash drive best? I could definitely spring for one big enough to hold my db if it'd help a lot, but it has that slower read speed. Does the faster seek time really make that much of a difference? Any brands you'd recommend? Neo4j stores the data as Graph on HD. An example: e = (n1,n2) e at location 1000 n1 at location 1 n2 at location 5 A traversal, assuming nothing is cached, would result in moving the head to 1 then to 1000 then back to 5. Normal HD take a while to move to the locations before it can start to read data. SSD does not have these delays. If you read little data that is spread widely over the storage, like in a traversal, SSD are much faster then HD even if they are slower to retrieve the data. I don't have performance data on that myself but I heard rumors of around 20-40 times speedup. cheers Martin On Thu, Aug 5, 2010 at 9:02 PM, Jeff Klann jkl...@iupui.edu wrote: Thanks for the answers. Yes, I can do online updates of the datastore, but while this is in RD I will need to rerun the main loop when I change the algorithm and just for personal benefit I don't want to wait hours to see the changes. Seems to be running acceptably now, though. However, I haven't benchmarked it against doing JOINS in Postgres. Are there any good performance stats out there? The speed is about the same as I'd expect from SQL. The graph will probably be nearly a complete graph in the end. The edges between orders will eventually store various stats on the relationships between pairs of items. It'd be nice if I can query an index for outgoing edges from nodes with certain properties. Is this possible? I'll have a look at the edge indexer component. Thanks, Jeff Klann On Mon, Aug 2, 2010 at 2:40 PM, David Montag david.mon...@neotechnology.com wrote: Hi Jeff, Please see answers below. On Mon, Aug 2, 2010 at 5:47 PM, Jeff Klann jkl...@iupui.edu wrote: Thank you all for your continued interest in helping me. I tweaked the code more to minimize writes to the database and it now looks like: For each item A For each customer that purchased A For each item B (with idA) that A purchased Increment (in memory) the weight of (A-B) Write out the edges [(A-B):weight] to disk and clear the in-memory map This actually (if I'm not mistaken) covers all relationships and does 7500 items in about 45 minutes! Not too bad, especially due to (I think) avoiding edge-checking, and I think it's usable for my application, though it's still ~200k traversals/sec on average, which is a few times slower than I hoped. I doubt that's much faster than a two-table join in SQL, though deeper traversals should show benefits. Do you need to do this computation on the full graph all the time? Maybe it would be enough to do it once, and then update it when a customer buys something? Usually, high one-time costs can be tolerated, and with Neo4j you can actually do the updating for a customer performing a purchase at runtime without performance problems. - David, thank you for your answers on traversers vs. getRelationships and on property-loading. I imported some properties I don't really need, perhaps if I delete them it'll speed things up? Also I'm using the old Node.traverse(). How is the new framework better? I expect it has a nicer syntax, which I would like to try, but does it improve performance too? Well, depending on your setup you should be able to theoretically improve performance compared to the old traversal framework. The old framework keeps track of visited nodes, so that you don't traverse to the same node twice. This behavior is customizable in the new framework. Please see http://wiki.neo4j.org/content/Traversal_Framework and check the Uniqueness constraints. If you know exactly when to stop, then you should be able to use Uniqueness.NONE, meaning that the framework does not keep track of visited nodes, meaning that you could end up traversing in a cycle. In your network however, you might know that you always traverse (item) --BOUGHT-- (customer) --BOUGHT-- (item) --CORRELATION-- (item)* and no further than that, so then you know that you won't end up in a cycle. But yeah, then you need to programmatically make sure you don't go too far. And I don't know if this gives you any performance benefits what so ever. Also, as I understand it, all properties for a node are loaded when they are first touched. Then they're kept in memory, so if you update properties later on the same node, and
Re: [Neo4j] Stumped by performance issue in traversal - would take a month to run!
Hi, there are some environmental optimizations you can do to speed things up. Neo4j is stored as a graph on disk, so traversal translate to moving the cursor on the hard drive if the data was not in RAM. For good performance you need a fast hd (flash drive would do best). Deleting lots of nodes can create holes in the db, so read operations have to move longer physical distance on the had drive then necessary. Only way I am aware of to get rid of holes reliably is to copy the DB into a fresh clean Neo4j DB. cheers Martin On Fri, Jul 30, 2010 at 8:10 PM, Jeff Klann jkl...@iupui.edu wrote: Hi, so I got 2GB more RAM and noticed that after adding some more memory map and increasing the heap space, my small query went from 6hrs to 3min. Quite reasonable! But the larger one that would take a month would still take a month. So I've been performance testing parts of it: The algorithm as in my first post showed *no* performance improvement on more RAM. But individual parts - Traversing only (first three lines) was much speedier, but still seems slow. 1.5 million traversals (15 out of 7000 items) took 23sec. It shaves off a few seconds if I run this twice and time it the second time, or if I don't print any node properties as I traverse. (Does Neo4J load ALL the properties for a node if one is accessed?) Even with a double run and not reading node properties, it still takes 16sec, which would make traversal take two hours. I thought Neo4J was suppposed to do ~1m traversals/sec, this is doing about 100k. Why? (And in fact on the other query it was getting about 800,000 traversals/sec.) Is one of Traversers vs. getRelationship iterators faster when getting all relationships of a type at depth 1? - Searching for relationships between A B (but not writing to them) takes it from 20s to 91s. Yuck. Maybe edge indexing is the way to avoid that? - Incrementing a property on the root node for every A B takes it from 20s to 61s (57s if it's all in one transaction). THAT seems weird. I imagine it has something to do with logging changes? Any way that can be turned off for a particular property (like it could be marked 'volatile' during a transaction or something)? I'm much more hopeful with the extra RAM but it's still kind of slow. Suggestions? Thanks, Jeff Klann On Wed, Jul 28, 2010 at 11:20 AM, Jeff Klann jkl...@iupui.edu wrote: Hi, I have an algorithm running on my little server that is very very slow. It's a recommendation traversal (for all A and B in the catalog of items: for each item A, how many customers also purchased another item in the catalog B). It's processed 90 items in about 8 hours so far! Before I dive deeper into trying to figure out the performance problem, I thought I'd email the list to see if more experienced people have ideas. Some characteristics of my datastore: it's size is pretty moderate for a database application. 7500 items, not sure how many customers and purchases (how can I find the size of an index?) but probably ~1 million customers. The relationshipstore + nodestore 500mb. (Propertystore is huge but I don't access it much in traversals.) The possibilities I see are: 1) *Neo4J is just slow.* Probably not slower than Postgres which I was using previously, but maybe I need to switch to a distributed map-reduce db in the cloud and give up the very nice graph modeling approach? I didn't think this would be a problem, because my data size is pretty moderate and Neo4J is supposed to be fast. 2) *I just need more RAM.* I definitely need more RAM - I have a measly 1GB currently. But would this get my 20day traversal down to a few hours? Doesn't seem like it'd have THAT much impact. I'm running Linux and nothing much else besides Neo4j, so I've got 650m physical RAM. Using 300m heap, about 300m memory-map. 3) *There's some secret about Neo4J performance I don't know.* Is there something I'm unaware that Neo4J is doing? When I access a property, does it load a chunk of properties I don't care about? For the current node/edge or others? I turned off log rotation and I commit after each item A. Are there other performance tips I might have missed? 4) *My algorithm is inefficient.* It's a fairly naive algorithm and maybe there's some optimizations I can do. It looks like: For each item A in the catalog: For each customer C that has purchased that item: For each item B that customer purchased: Update the co-occurrence edge between AB. (If the edge exists, add one to its weight. If it doesn't exist, create it with weight one.) This is O(n^2) worst case, but practically it'll be much better due to the sparseness of purchases. The large number of customers slows it down, though. The slowest part, I suspect, is the last line. It's a lot of finding and re-finding edges between As and Bs and updating the edge properties. I
Re: [Neo4j] Stumped by performance issue in traversal - would take a month to run!
Hi, there are some environmental optimizations you can do to speed things up. Neo4j is stored as a graph on disk, so traversal translate to moving the cursor on the hard drive if the data was not in RAM. For good performance you need a fast hd (flash drive would do best). Deleting lots of nodes can create holes in the db, so read operations have to move longer physical distance on the had drive then necessary. Only way I am aware of to get rid of holes reliably is to copy the DB into a fresh clean Neo4j DB. cheers Martin On Fri, Jul 30, 2010 at 8:10 PM, Jeff Klann jkl...@iupui.edu wrote: Hi, so I got 2GB more RAM and noticed that after adding some more memory map and increasing the heap space, my small query went from 6hrs to 3min. Quite reasonable! But the larger one that would take a month would still take a month. So I've been performance testing parts of it: The algorithm as in my first post showed *no* performance improvement on more RAM. But individual parts - Traversing only (first three lines) was much speedier, but still seems slow. 1.5 million traversals (15 out of 7000 items) took 23sec. It shaves off a few seconds if I run this twice and time it the second time, or if I don't print any node properties as I traverse. (Does Neo4J load ALL the properties for a node if one is accessed?) Even with a double run and not reading node properties, it still takes 16sec, which would make traversal take two hours. I thought Neo4J was suppposed to do ~1m traversals/sec, this is doing about 100k. Why? (And in fact on the other query it was getting about 800,000 traversals/sec.) Is one of Traversers vs. getRelationship iterators faster when getting all relationships of a type at depth 1? - Searching for relationships between A B (but not writing to them) takes it from 20s to 91s. Yuck. Maybe edge indexing is the way to avoid that? - Incrementing a property on the root node for every A B takes it from 20s to 61s (57s if it's all in one transaction). THAT seems weird. I imagine it has something to do with logging changes? Any way that can be turned off for a particular property (like it could be marked 'volatile' during a transaction or something)? I'm much more hopeful with the extra RAM but it's still kind of slow. Suggestions? Thanks, Jeff Klann On Wed, Jul 28, 2010 at 11:20 AM, Jeff Klann jkl...@iupui.edu wrote: Hi, I have an algorithm running on my little server that is very very slow. It's a recommendation traversal (for all A and B in the catalog of items: for each item A, how many customers also purchased another item in the catalog B). It's processed 90 items in about 8 hours so far! Before I dive deeper into trying to figure out the performance problem, I thought I'd email the list to see if more experienced people have ideas. Some characteristics of my datastore: it's size is pretty moderate for a database application. 7500 items, not sure how many customers and purchases (how can I find the size of an index?) but probably ~1 million customers. The relationshipstore + nodestore 500mb. (Propertystore is huge but I don't access it much in traversals.) The possibilities I see are: 1) *Neo4J is just slow.* Probably not slower than Postgres which I was using previously, but maybe I need to switch to a distributed map-reduce db in the cloud and give up the very nice graph modeling approach? I didn't think this would be a problem, because my data size is pretty moderate and Neo4J is supposed to be fast. 2) *I just need more RAM.* I definitely need more RAM - I have a measly 1GB currently. But would this get my 20day traversal down to a few hours? Doesn't seem like it'd have THAT much impact. I'm running Linux and nothing much else besides Neo4j, so I've got 650m physical RAM. Using 300m heap, about 300m memory-map. 3) *There's some secret about Neo4J performance I don't know.* Is there something I'm unaware that Neo4J is doing? When I access a property, does it load a chunk of properties I don't care about? For the current node/edge or others? I turned off log rotation and I commit after each item A. Are there other performance tips I might have missed? 4) *My algorithm is inefficient.* It's a fairly naive algorithm and maybe there's some optimizations I can do. It looks like: For each item A in the catalog: For each customer C that has purchased that item: For each item B that customer purchased: Update the co-occurrence edge between AB. (If the edge exists, add one to its weight. If it doesn't exist, create it with weight one.) This is O(n^2) worst case, but practically it'll be much better due to the sparseness of purchases. The large number of customers slows it down, though. The slowest part, I suspect, is the last line. It's a lot of finding and re-finding edges between As and Bs and updating the edge properties. I
Re: [Neo4j] memory consumption
Do you use the Batchinserter or a normal Transaction? When using a normal Transaction to insert huge amounts of data I always submit and create a new transaction every X Items. This keeps the transaction small and reduces the memory used. cheers Martin On Sun, Jul 4, 2010 at 4:13 PM, qiuyan...@mailbox.tu-berlin.de wrote: Hallo, I'm currently working with neo4j database and want to insert a bunch of data into it. At the very beginning the program works quite well. But as more data has been inserted into the database, the insertion runs more and more slowly and I noticed that the program consumes really a lot of memory. Even though I splitted the input file into small pieces so that each time the program tries only to insert a small part of data, the problem occurs. That means, as there exists already much data in the database, the program consumes a lot of memory as soon as it begins so that the insertion is so slow that it seems that it won't be able to finish. I wonder if there's some solutions to save the memory. Thanks in advance. Cheers, Qiuyan ___ Neo4j mailing list User@lists.neo4j.org https://lists.neo4j.org/mailman/listinfo/user ___ Neo4j mailing list User@lists.neo4j.org https://lists.neo4j.org/mailman/listinfo/user
[Neo4j] Tell neo to not reuse ID's
Hej, Is it somehow possible to tell Neo4j not to reuse id's at all? Im running some experiments on Neo4j and I want to add and delete the nodes and relationships. To make sure that I can repeat the same experiment I create a log containing the ID's of the nodes i want to delete. To make sure that I can rerun the experiment each node I add has to have the same ID in each experiment. If ID's can be reused that is not always the case thats why I need to turn it off or work around it. hope for your help cheers Martin ___ Neo4j mailing list User@lists.neo4j.org https://lists.neo4j.org/mailman/listinfo/user