There is nothing crazy with async I/O. Netty (which I presume is underlying network library in a latest Hadoop) is totally async. You can run 1000, 10,000 and may be more threads on Linux but performance -wise this would sub-optimal decision. With 20 ms time share, on 8 core CPU some threads can wait their schedule slot for up to 2.5 seconds (1000 threads) or 25 sec (10000 secs) or 250 secs (100,000). This is worst case scenario, of course. If you do not care about latency it probably won't hurt you too much. Another problems with a large number of threads scheduling
1. is time which kernel scheduler spends doing nothing (from the application point of view). 2. L1/L2/L3 cache trashing One more: having total number of active threads less than a number of physical cores (or threads in SPARC TX) is #1 requirement of a high-performance application. This allows you to cut on threads synchronization cost by utilizing such expensive things like spin locks (busy loops). Stack has already mentioned LMAX Disruptor framework which is totally bypasses standard Java thread synchronizations (they use CAS, busy loops and thread yielding). Best regards, Vladimir Rodionov Principal Platform Engineer Carrier IQ, www.carrieriq.com e-mail: vrodio...@carrieriq.com ________________________________________ From: lars hofhansl [lhofha...@yahoo.com] Sent: Monday, October 24, 2011 12:04 PM To: dev@hbase.apache.org Subject: Re: SILT - nice keyvalue store paper I am not sure I would generally agree with this statement. The linux kernel can easily handle 10.000's of threads (just need to keep default stack small). (there were tests done with 1m threads too). Doing async IO is all the craze today (see node.js and friends), but that also is not necessarily done with full understanding of the performance characteristics. Just my $0.02. ----- Original Message ----- From: Vladimir Rodionov <vrodio...@carrieriq.com> To: "dev@hbase.apache.org" <dev@hbase.apache.org> Cc: Sent: Monday, October 24, 2011 11:43 AM Subject: RE: SILT - nice keyvalue store paper Jonathan, 1000 threads is a bad application design. They will kill you even w/o contention. In my opinion, over-usage of a *modern* Java concurrent framework stuff w/o real understanding of a benefits is a major contributor to a poor performance and poor scalability of a Java application. Before making any decision on using a fancy data structure which can potentially affect application performance it is always a good idea to run some benchmarks. Best regards, Vladimir Rodionov Principal Platform Engineer Carrier IQ, www.carrieriq.com e-mail: vrodio...@carrieriq.com ________________________________________ From: Jonathan Gray [jg...@fb.com] Sent: Sunday, October 23, 2011 4:20 PM To: dev@hbase.apache.org Subject: RE: SILT - nice keyvalue store paper Very nice experiment, Akash. Keep getting your hands dirty and digging! :) I think your results might change if you bump the test up to 1000 threads or so. 100 threads can still perform okay when there's a global lock but the contention at 1000 threads will kill you and that's when CSLM should do much better. (1000 handler threads is approx. what I run with on RS in prod). Though I am a bit surprised that at 100 threads the TreeMap was significantly faster. Your inconsistent results are a bit odd, you might try an order of magnitude more operations per thread. You might also gather some statistics about tree size and per operation latency. I've done some isolated CSLM benchmarks in the past and have never been able to reproduce any of the slowness people suggest. I recall trying some impractically large MemStores and everything still being quite fast. Over in Cassandra, I believe they have a two-level CSLM with the first map key being the row and then the columns for each row in their own CSLM. I've been told this is somewhat of a pain point for them. And keep in mind they have one shard/region per node and we generally have several smaller MemStores on each node (tens to thousands). Not sure we would want to try that. There could be some interesting optimizations if you had very specific issues, like if you had a ton of reads to MemStore and not many writes you could keep some kind of mirrored hashmap. And for writes, the WAL is definitely the latency bottleneck. But if you are doing lots of small operations, so your WALEdits are not large, and with some of the HLog batching features going in to trunk, you end up with hundreds of requests per HLog sync. And although the syncs are higher latency, with batching you end up getting high throughput. And the bottleneck shifts. Each sync will take approx. 1-5ms, so let's say 250 requests per HLog sync batch, 4ms per sync, so 62.5k req/sec. (62.5k * 100 bytes/req = 600K/sec, very reasonable). If you're mixing in reads as well (or if you're doing increments which do a read and write), then this adds to the CPU usage and contention without adding to HLog throughput. All of a sudden the bottleneck becomes CPU/contention and not HLog latency or throughput. Highly concurrent increments/counters with a largely in-memory dataset can easily be CPU bottlenecked. For one specific application Dhruba and I worked on, we made some good improvements in CPU efficiency by reducing the number of operations and increasing efficiency on the CSLM. Doing things like always taking a tailMap and working from that instead of starting at the root node, using an iterator() and taking advantage of the available remove() semantics, or simply just mutating things that are normally immutable :) Unfortunately many of these optimizations were semi-horrid hacks and introduced things like ModifiableKeyValues, so they all haven't made their way to apache. In the end, after our optimizations, the real world workload Dhruba and I were working with was not all in-memory so the bottleneck in production became the random reads (so increasing the block cache hit ratio is the focus) rather than CPU contention or HLog throughput. JG From: Akash Ashok [mailto:thehellma...@gmail.com] Sent: Sunday, October 23, 2011 2:57 AM To: dev@hbase.apache.org Subject: Re: SILT - nice keyvalue store paper I was running some similar tests and came across a surprising finding. I compared reads and write on ConcurrentSkipListMap ( which the memstore uses) and synchronized TreeMap ( Which was literally treemap synchronized). Executed concurrent reads, writes and deletes on both of them. Surprisingly synchronized treeMap performed better, though just slightly better, than ConcurrentSkipListMap which KeyValueSkipListSet uses. Here are the output of a few runs Sometimes the difference was considerable Using HBaseMap it took 20438ms Using TreeMap it took 11613ms Time Difference:8825ms And sometimes the difference was negligible Using HBaseMap it took 13370ms Using TreeMap it took 9482ms Time Difference:3888ms I've attaching the test java file which I wrote to test it. This might be a very minor differece but still surprising considering the fact that ConcurrentSkipListMap uses fancy 2 level indexes which they say improves the deletion performance. And here are the details about the test run. 100 Threads each fetching 1,000,000 records 100 threads each adding 1,000,000 records. 100 threads each deletin 1,000,000 records ( Reads, Writes and deletes simultaneously ) Cheers, Akash A On Sun, Oct 23, 2011 at 3:25 AM, Stack <st...@duboce.net<mailto:st...@duboce.net>> wrote: On Sat, Oct 22, 2011 at 2:41 PM, N Keywal <nkey...@gmail.com<mailto:nkey...@gmail.com>> wrote: > I would think that the bottleneck for insert is the wal part? > It would be possible to do a kind of memory list preparation during the wal > insertion, and if the wal insertion is confirmed, do the insertion in the > memory list. But it's strange to have the insertion in memory important vs. > the insertion on disk... > Yes, WAL is the long pole writing. But MemStore has issues too; Dhruba says cpu above. Reading and writing it is also 'slow'. St.Ack
