There’s two parts to latency on the Cassandra side:
Local and coordinator
When you read, the node to which you connect coordinates the request to the
node which has the data (potentially itself). Long tail in coordinator
latencies tend to be the coordinator itself gc’ing, which will happen from time
to time. If it’s more consistently high, it may be natural latencies in your
cluster (ie: your requests are going cross wan and the other dc is 10-20ms
away).
If the latency is seen in p99 but not p50, you can almost always speculatively
read from another coordinator (driver level speculation) after a millisecond or
so.
--
Jeff Jirsa
> On Feb 5, 2018, at 5:41 AM, mohsin k <moshinkarova...@gmail.com> wrote:
>
> Thanks for response @Nicolas. I was considering the total read latency from
> the client to server (as shown in the image above) which is around 30ms.
> Which I want to get around 3ms (client and server are both on same network).
> I did not consider read latency provided by the server (which I should have).
> I monitored CPU , memory and JVM lifecycle, which is at a safe level. I think
> the difference(0.03 to 30) might be because of low network bandwidth, correct
> me if I am wrong.
>
> I did reduce chunk_length_in_kb to 4kb, but I couldn't get a considerable
> amount of difference, might be because there is less room for improvement on
> the server side.
>
> Thanks again.
>
>> On Mon, Feb 5, 2018 at 6:52 PM, Nicolas Guyomar <nicolas.guyo...@gmail.com>
>> wrote:
>> Your row hit rate is 0.971 which is already very high, IMHO there is
>> "nothing" left to do here if you can afford to store your entire dataset in
>> memory
>>
>> Local read latency: 0.030 ms already seems good to me, what makes you think
>> that you can achieve more with the relative "small" box you are using ?
>>
>> You have to keep an eye on other metrics which might be a limiting factor,
>> like cpu usage, JVM heap lifecycle and so on
>>
>> For read heavy workflow it is sometimes advised to reduce chunk_length_in_kb
>> from the default 64kb to 4kb, see if it helps !
>>
>>> On 5 February 2018 at 13:09, mohsin k <moshinkarova...@gmail.com> wrote:
>>> Hey Rahul,
>>>
>>> Each partition has around 10 cluster keys. Based on nodetool, I can roughly
>>> estimate partition size to be less than 1KB.
>>>
>>>> On Mon, Feb 5, 2018 at 5:37 PM, mohsin k <moshinkarova...@gmail.com> wrote:
>>>> Hey Nicolas,
>>>>
>>>> My goal is to reduce latency as much as possible. I did wait for warmup.
>>>> The test ran for more than 15mins, I am not sure why it shows 2mins though.
>>>>
>>>>
>>>>
>>>>> On Mon, Feb 5, 2018 at 5:25 PM, Rahul Singh
>>>>> <rahul.xavier.si...@gmail.com> wrote:
>>>>> What is the average size of your partitions / rows. 1GB may not be enough.
>>>>>
>>>>> Rahul
>>>>>
>>>>>> On Feb 5, 2018, 6:52 AM -0500, mohsin k <moshinkarova...@gmail.com>,
>>>>>> wrote:
>>>>>> Hi,
>>>>>>
>>>>>> I have been looking into different configurations for tuning my
>>>>>> cassandra servers. So, initially I loadtested server using
>>>>>> cassandra-stress tool, with default configs and then tuning one by one
>>>>>> config to measure impact of change. First config, I tried was setting
>>>>>> "row_cache_size_in_mb" to 1000 (MB) in yaml, adding caching {'keys':
>>>>>> 'ALL', 'rows_per_partition': 'ALL'}. After changing these configs, I
>>>>>> observed that latency has increased rather than decreasing. It would be
>>>>>> really helpful if I get to understand why is this the case and what
>>>>>> steps must be taken to decrease the latency.
>>>>>>
>>>>>> I am running a cluster with 4 nodes.
>>>>>>
>>>>>> Following is my schema:
>>>>>>
>>>>>> CREATE TABLE stresstest.user_to_segment (
>>>>>> userid text,
>>>>>> segmentid text,
>>>>>> PRIMARY KEY (userid, segmentid)
>>>>>> ) WITH CLUSTERING ORDER BY (segmentid DESC)
>>>>>> AND bloom_filter_fp_chance = 0.1
>>>>>> AND caching = {'keys': 'ALL', 'rows_per_partition': 'ALL'}
>>>>>> AND comment = 'A table to hold blog segment user relation'
>>>>>> AND compaction = {'class':
>>>>>> 'org.apache.cassandra.db.compaction.LeveledCompactionStrategy'}
>>>>>> AND compression = {'chunk_length_in_kb': '64', 'class':
>>>>>> 'org.apache.cassandra.io.compress.LZ4Compressor'}
>>>>>> AND crc_check_chance = 1.0
>>>>>> AND dclocal_read_repair_chance = 0.1
>>>>>> AND default_time_to_live = 0
>>>>>> AND gc_grace_seconds = 864000
>>>>>> AND max_index_interval = 2048
>>>>>> AND memtable_flush_period_in_ms = 0
>>>>>> AND min_index_interval = 128
>>>>>> AND read_repair_chance = 0.0
>>>>>> AND speculative_retry = '99PERCENTILE';
>>>>>>
>>>>>> Following are node specs:
>>>>>> RAM: 4GB
>>>>>> CPU: 4 Core
>>>>>> HDD: 250BG
>>>>>>
>>>>>>
>>>>>> Following is the output of 'nodetool info' after setting
>>>>>> row_cache_size_in_mb:
>>>>>>
>>>>>> ID : d97dfbbf-1dc3-4d95-a1d9-c9a8d22a3d32
>>>>>> Gossip active : true
>>>>>> Thrift active : false
>>>>>> Native Transport active: true
>>>>>> Load : 10.94 MiB
>>>>>> Generation No : 1517571163
>>>>>> Uptime (seconds) : 9169
>>>>>> Heap Memory (MB) : 136.01 / 3932.00
>>>>>> Off Heap Memory (MB) : 0.10
>>>>>> Data Center : dc1
>>>>>> Rack : rack1
>>>>>> Exceptions : 0
>>>>>> Key Cache : entries 125881, size 9.6 MiB, capacity 100 MiB,
>>>>>> 107 hits, 126004 requests, 0.001 recent hit rate, 14400 save period in
>>>>>> seconds
>>>>>> Row Cache : entries 125861, size 31.54 MiB, capacity 1000
>>>>>> MiB, 4262684 hits, 4388545 requests, 0.971 recent hit rate, 0 save
>>>>>> period in seconds
>>>>>> Counter Cache : entries 0, size 0 bytes, capacity 50 MiB, 0
>>>>>> hits, 0 requests, NaN recent hit rate, 7200 save period in seconds
>>>>>> Chunk Cache : entries 273, size 17.06 MiB, capacity 480 MiB,
>>>>>> 325 misses, 126623 requests, 0.997 recent hit rate, NaN microseconds
>>>>>> miss latency
>>>>>> Percent Repaired : 100.0%
>>>>>> Token : (invoke with -T/--tokens to see all 256 tokens)
>>>>>>
>>>>>>
>>>>>> Following is output of nodetool cfstats:
>>>>>>
>>>>>> Total number of tables: 37
>>>>>> ----------------
>>>>>> Keyspace : stresstest
>>>>>> Read Count: 4398162
>>>>>> Read Latency: 0.02184742626579012 ms.
>>>>>> Write Count: 0
>>>>>> Write Latency: NaN ms.
>>>>>> Pending Flushes: 0
>>>>>> Table: user_to_segment
>>>>>> SSTable count: 1
>>>>>> SSTables in each level: [1, 0, 0, 0, 0, 0, 0, 0, 0]
>>>>>> Space used (live): 11076103
>>>>>> Space used (total): 11076103
>>>>>> Space used by snapshots (total): 0
>>>>>> Off heap memory used (total): 107981
>>>>>> SSTable Compression Ratio: 0.5123353861375962
>>>>>> Number of partitions (estimate): 125782
>>>>>> Memtable cell count: 0
>>>>>> Memtable data size: 0
>>>>>> Memtable off heap memory used: 0
>>>>>> Memtable switch count: 2
>>>>>> Local read count: 4398162
>>>>>> Local read latency: 0.030 ms
>>>>>> Local write count: 0
>>>>>> Local write latency: NaN ms
>>>>>> Pending flushes: 0
>>>>>> Percent repaired: 0.0
>>>>>> Bloom filter false positives: 0
>>>>>> Bloom filter false ratio: 0.00000
>>>>>> Bloom filter space used: 79280
>>>>>> Bloom filter off heap memory used: 79272
>>>>>> Index summary off heap memory used: 26757
>>>>>> Compression metadata off heap memory used: 1952
>>>>>> Compacted partition minimum bytes: 43
>>>>>> Compacted partition maximum bytes: 215
>>>>>> Compacted partition mean bytes: 136
>>>>>> Average live cells per slice (last five minutes): 5.719932432432432
>>>>>> Maximum live cells per slice (last five minutes): 10
>>>>>> Average tombstones per slice (last five minutes): 1.0
>>>>>> Maximum tombstones per slice (last five minutes): 1
>>>>>> Dropped Mutations: 0
>>>>>>
>>>>>> Following are my results:
>>>>>> The blue graph is before setting row_cache_size_in_mb,
>>>>>> orange is after
>>>>>>
>>>>>> Thanks,
>>>>>> Mohsin
>>>>>>
>>>>>>
>>>>>> ---------------------------------------------------------------------
>>>>>> To unsubscribe, e-mail: user-unsubscr...@cassandra.apache.org
>>>>>> For additional commands, e-mail: user-h...@cassandra.apache.org
>>>>
>>>
>>
>
