Tuesday, November 4, 2014


Assumption and purpose:
This is an attempt to compare two promising open source Object Store Technologies purely based on performance. The use case kept in mind is large or small scale public cloud storage provider & the attempt here is evaluate the best technology for said use case.
Feature delta between OpenStack Swift and Ceph Object Store is ignored here.  Ceph is viewed only as Object Store serving Objects via Swift REST API (not RADOS Objects), Ceph’s other interfaces which provide file and block based access are ignored here.    
Assumption here is both the technologies can be best compared when deployed on same hardware and topology  & tested with same kind of workload. Data caching is discouraged while collecting numbers (Page Cache, Dentries and Inode are flushed every minute on each server). COSBench is used as benchmarking tool.


I got some suggestion to improve Ceph-RGW performance from Ceph community . I tried all of them, they do have some minor impact on the overall Ceph-RGW performance(<3%). However there is nothing that change the overall conclusion of the study.

It would not be called an apple to apple comparison but with multiple RGW-civetweb instances and HA proxy, I was able to get better results with Ceph-RGW. I will be posting them soon.


There are two flavors of Dell Power Edge R620 servers used in the study. For simplicity I will now call them T1 & T2.
CPU: 2x Intel-E5-2680 10C 2.8GHz 25M$ (40 Logical CPU, with HT Enabled)
RAM: 4x 16GB RDIMM, dual rank x4 (64GB)
NIC1: Broadcom Corporation NetXtreme II BCM57810 10 Gigabit Ethernet (For Management)
NIC2: Mellonix Connect-X3, 40 Gigabit Ethernet, Dual Port Full Duplex (For Data)
Storage:  160 GB HDD (For OS).
CPU: 2x Intel-E5-2680 10C 2.8GHz 25M$ (40 Logical CPU, with HT Enabled)
RAM: 8x 16GB RDIMM, dual rank x4 (128GB)
NIC1: Broadcom Corporation NetXtreme II BCM57810 10 Gigabit Ethernet (For Management)
NIC2: Mellonix Connect-X3, 40 Gigabit Ethernet, Dual Port Full Duplex (For Data)
Storage1: 160 GB HDD (For OS)
Storage2:  10x 400GB Optimus Eco™ 2.5” SAS SSDs (4TB)
Interface SAS (4 Phy 6Gb/s)
Interface Ports Dual/Wide

Network Bandwidth Check:

Host-A$ date ; sudo iperf -c XXX.XXX.XXX.B  -p 5001 -P4 -m ; date
Tue Nov  4 13:16:58 IST 2014
Client connecting to XXX.XXX.XXX.B  , TCP port 5001
TCP window size:  325 KByte (default)
[  5] local XXX.XXX.XXX.A  port 43892 connected with XXX.XXX.XXX.B  port 5001
[  3] local XXX.XXX.XXX.A  port 43891 connected with XXX.XXX.XXX.B  port 5001
[  6] local XXX.XXX.XXX.A  port 43893 connected with XXX.XXX.XXX.B  port 5001
[  4] local XXX.XXX.XXX.A  port 43890 connected with XXX.XXX.XXX.B  port 5001
[ ID] Interval       Transfer     Bandwidth
[  5]  0.0-10.0 sec  10.9 GBytes  9.35 Gbits/sec
[  5] MSS size 8960 bytes (MTU 9000 bytes, unknown interface)
[  3]  0.0-10.0 sec  9.17 GBytes  7.88 Gbits/sec
[  3] MSS size 8960 bytes (MTU 9000 bytes, unknown interface)
[  6]  0.0-10.0 sec  16.5 GBytes  14.2 Gbits/sec
[  6] MSS size 8960 bytes (MTU 9000 bytes, unknown interface)
[  4]  0.0-10.0 sec  8.72 GBytes  7.49 Gbits/sec
[  4] MSS size 8960 bytes (MTU 9000 bytes, unknown interface)
[SUM]  0.0-10.0 sec  45.3 GBytes  38.9 Gbits/sec
Tue Nov  4 13:17:08 IST 2014
Host-B$ date ; sudo iperf -c -p 4001 -P4 -m ; date
Tue Nov  4 13:17:01 IST 2014
Client connecting to, TCP port 4001
TCP window size:  325 KByte (default)
[  4] local XXX.XXX.XXX.B port 59130 connected with XXX.XXX.XXX.A port 4001
[  3] local XXX.XXX.XXX.B port 59131 connected with XXX.XXX.XXX.A port 4001
[  6] local XXX.XXX.XXX.B port 59133 connected with XXX.XXX.XXX.A port 4001
[  5] local XXX.XXX.XXX.B port 59132 connected with XXX.XXX.XXX.A port 4001
[ ID] Interval       Transfer     Bandwidth
[  4]  0.0-10.0 sec  14.6 GBytes  12.6 Gbits/sec
[  4] MSS size 8960 bytes (MTU 9000 bytes, unknown interface)
[  3]  0.0-10.0 sec  7.90 GBytes  6.79 Gbits/sec
[  3] MSS size 8960 bytes (MTU 9000 bytes, unknown interface)
[  6]  0.0-10.0 sec  14.7 GBytes  12.7 Gbits/sec
[  6] MSS size 8960 bytes (MTU 9000 bytes, unknown interface)
[  5]  0.0-10.0 sec  8.40 GBytes  7.21 Gbits/sec
[  5] MSS size 8960 bytes (MTU 9000 bytes, unknown interface)
[SUM]  0.0-10.0 sec  45.7 GBytes  39.2 Gbits/sec
Tue Nov  4 13:17:11 IST 2014

So the total available bandwidth is ~39Gbps(~5GBps) for inbound and ~39Gbps(~5GBps) for  outbound traffic as well.

Topology & Setup:

Ceph setup has two more monitor nodes which are not show here.

Ceph RGW Setup:

OpenStack Swift Setup:

Software Details

General Configuration:

  1. Ubuntu 14.04 (3.13.0-24-generic)
  2. Linux Tuning options for networking configured on all the nodes
#Configs recommended for Mellonix Connect –X3
net.ipv4.tcp_timestamps = 0
net.ipv4.tcp_sack = 0
net.ipv4.tcp_low_latency = 1
net.core.netdev_max_backlog = 250000
net.core.rmem_max = 16777216
net.core.wmem_max = 16777216
net.core.rmem_default = 16777216
net.core.wmem_default = 16777216
net.core.optmem_max = 16777216
net.ipv4.tcp_rmem = 4096 87380 16777216
net.ipv4.tcp_wmem = 4096 65536 16777216
net.ipv4.tcp_tw_reuse = 1
net.ipv4.tcp_tw_recycle = 1
kernel.core_uses_pid = 1

MTU size of 9000 is used along with above options.
  1. A CRON job is configured to flush DRAM cache each minute on each node.

sudo sh -c "sync; echo 3 > /proc/sys/vm/drop_caches"

Ceph Configurations:-

  1. Ceph Version: 0.87
  2. RGW is used with Apache + FASTCGI as well as CivetWeb.
  3. Apache version : 2.4.7-1ubuntu4.1, with libapache2-mod-fastcgi -2.4.7~0910052141-1.1
  4. Ceph.conf is placed here. This contains the entire ceph optimization configuration, done in the experiment.  
  5. Default region, zone and pools. All .rgw *pools created with default zone are set to use PG_NUM of 4096.
  6. Replica count is set to 3.(Max_Size=3, Min_Size=2)
  7. Apache configuration parameters:
ServerLimit          4096
ThreadLimit           200
StartServers           20
MinSpareThreads        30
MaxSpareThreads       100
ThreadsPerChild       128
MaxClients           4096
MaxRequestsPerChild  10000  
  1. CivetWeb is used with all the default configurations. However ‘rgw_op_thread’ seems to control the CivetWeb’s configuration option ‘num_op_thread’, which is set to 128. This parameter seems to degrade the performance in term of response time, if increased beyond this point. (I tried setting this to 256/512 and it resulted in to CLOSE_WAIT state of more & more HTTP connections). I am hitting a CivetWeb bug related to this problem.   

OpenStack Swift Configurations:

  1. OpenStack Swift Version :  Icehouse/ Swift2.0
  2. Webserver: Default WSGI
  3. All performance optimization is done based on OpenStack Swift deployment guide.
  4. Inode Size of 256K is used all other XFS formatting and mounting options are as per recommendation made in Swift Deployment Guide.
  5. WSGI pipeline is trimmed down and only contains essential middleware.
Proxy Server WSGI pipeline looks like this:
pipeline = healthcheck cache tempauth proxy-server   
  1. Each Storage node is configured as a zone in a region, and on each node there is disk that is dedicatedly used for Account and Container Databases.  All the other disks are used for keeping Objects only. Ring files are populated based on these configurations.
  2. Proxy node only runs the proxy-server & memcached.
  3. Storage node run all the other swift services i.e account-server, container-server, object-server along with supporting services  like auditors, updaters, replicaters.

COSBench & Workload Details

  1. COSBench Version: 0.4.0.e1  
  2. COSBench Controller and driver both are configured on the same machine, as the hardware is capable of sustaining the workload.
  3. Small File/Object workload is as follows:
Object Size: 1MB
Containers: 100
Objects Per Container: 1000
  1. Large File/Object workload is as follows:
Object Size: 1GB
Containers: 10
Objects Per Container: 100

  1. Objects are written once in both the cases.
  2. Every workload is configured to use different COSBench worker count.
  3. For Small File Workload Worker counts are: 32,64,128,256,512
  4. For Large File Workload Worker counts  are: 8,16,32,64,128
  5. Each Workload is executed for 900 Seconds, and objects are read randomly from the available set of Swift Objects.
  6. There is no difference in workloads for Ceph and Swift except the value of generated token. A token was generated after creating Swift users in both cases. This token is provided along with Storage-URL in workload configurations.
  7. Ceph Put all the Swift objects in Single Ceph Pool called ‘.rgw.buckets’.


Small File Workload:

Large File Workload:

Additional Details:

90%RT: Response Time of 90% requests
Max RT: Maximum Response Time taken by all successful requests.  

Small Files

Worker Count
90%RT< 20 ms
Max RT=1450 ms
90%RT< 20 ms
Max RT=1440 ms
90%RT< 30 ms
Max RT=10,230 ms
90%RT< 50 ms
Max RT=2000 ms
90%RT< 60 ms
Max RT= 1,460 ms
90%RT< 30 ms
Max RT=16,336 ms
90%RT< 110 ms
Max RT=3090 ms
90%RT< 120 ms
Max RT=  1,480 ms
90%RT< 70 ms
Max RT=16,380ms
90%RT< 210 ms
Max RT=1760 ms
90%RT< 120 ms
Max RT=  3280 ms
90%RT< 90 ms
Max RT=17,020ms
90%RT< 330 ms
Max RT=33,120 ms
90%RT< 200 ms
Max RT=  20,040 ms
90%RT< 160 ms
Max RT=16,760ms

Large Files:

Worker Count
90%RT< 3,110 ms
Max RT= 4,540 ms
90%RT < 3,060 ms
Max RT= 5,380 ms
90%RT < 6,740 ms
Max RT= 11,210 ms
90%RT< 5,550 ms
Max RT= 7,980  ms
90%RT < 5,780 ms
Max RT= 18,150 ms
90%RT < 8,150 ms
Max RT= 13,710 ms
90%RT< 10,860 ms
Max RT= 11,900 ms
90%RT < 10,970 ms
Max RT= 12,120 ms
90%RT < 9,800 ms
Max RT= 17,810 ms
90%RT< 21,370 ms
Max RT=  24,200 ms
90%RT < 21,190 ms
Max RT= 22,080 ms
90%RT < 19,530 ms
Max RT= 38,760 ms
90%RT< 42,410 ms
Max RT= 43,340 ms
90%RT < 41,590 ms
Max RT= 44,210 ms
90%RT < 46,800 ms
Max RT=74,810 ms


  1. Native Swift behaviour & results curve seems sane. A clear relation between concurrency and throughput is established. 
  2. Ceph-RGW seems to have problem with RGW threading model, a flat throughput curve with increased concurrency is certainly not a good sign.
  3. Native Swift in general performs better in high concurrency environments.
  4. Ceph RGW gives better bandwidth at lower concurrency.
  5. Ceph RGW response time is excellent for Large Objects.
  6. For Small Objects at lower concurrency, Ceph -RGW seems very promising, however there is much to do, as concurrency plays a great role in Web Server environment.
  7. Ceph RGW major bottleneck is WebServer, however CivetWeb  & Apache FASCGI gives comparable numbers. However CivetWeb is better than apache+fcgi in term of response time at high concurrency. CivetWeb has a inherent design limitation, which is already reported here.  
  8. Digging further I also made an attempt to benchmark Ceph using RADOS bench, which directly uses Ceph Objects(Different from Swift Object interface it provides). I ran the bench the same node which is used as COSBench Controller +Driver. So in this case RGW is out of the picture , in summary my observations are as below:
Object Size& Threads
Avg Bandwidth (MB/Sec)
Avg Latency (Sec)
Runtime (Sec)
1M , t=128
1M, t=256
4M, t=128

Bandwidth number can be directly used for representing IOPs. So in summary even RADOS bench is not giving bandwidth beyond ~4GB/s. Strange thing here is it is optimize for 4MB Size, increasing beyond this Object size is not giving higher OPs (bandwidth).

Other Remarks:
  1. Swift is more feature rich in terms of REST API.
  2. S3 API is supported by both.
  3. Finding good documentation is big pain in setting up Ceph.


  1. Good doc Pushpesh :)

  2. I am posting some of the mail conversation, that might be useful for everyone.

  3. Ben England
    Nov 5 (1 day ago)

    to me, Mark, Thiago, Peter, Kyle, John, Neil
    Thanks Pushpesh, this should give us a lot of ideas about how to tune Ceph and RGW in particular. I'm having some trouble understanding how to configure it so it will help there too. cc'ing Kyle Bader (RGW developer) in case this interests him.



  4. Mark Nelson
    Nov 5 (1 day ago)

    to Sage, yehuda, Ben, me, Thiago, Peter, Kyle, John, Neil
    Hi Pushpesh,

    A couple of comments:

    I noticed that you are doing:

    sudo sh -c "sync; echo 3 > /proc/sys/vm/drop_caches"

    every minute which will flush dentry/inode cache. With how ceph lays objects out on the disk (and I suspect swift too) this is will hurt performance for small objects. Given that you are using SSDs I'm not sure how much effect it will have, but it's something to keep in mind. Are you re-reading objects? If so, maybe an "echo 1" here might be better to avoid cached reads while keeping dentries/inodes in cache?

    Also, with the number of total objects in the 1MB object tests, you might want to tweak these parameters:


    By default those settings restrict each PG to 300 objects before the nested directory structure where objects are stored for each PG is split into a deeper tree. With 1 million objects, 3x replication, and 4096 PGs, you likely hit those limits and caused the directory structure to split at least once. With traditional spinning disks this can have a big performance impact, though again I'm not sure how much of one you would see with SSDs.

    Having said that, it's very interesting that your civitweb results for RGW are so much faster than the apache results. I'm not an apache tuning expert, so perhaps there is more tuning we can do on that side.

    If you'd like, we have a weekly public ceph performance meeting on Wednesday mornings at 8AM PST. We'd be happy to have you join and talk about what you've been seeing! In particular, Somnath Roy from Sandisk has been doing a ton of work optimizing the read path in Ceph and RGW and might have some additional comments on your tests.

    Here's the meeting etherpad:


    And the meeting URL:



  5. Yehuda Sadeh
    Nov 5 (1 day ago)

    to Mark, Ben, me, Thiago, Peter, Kyle, John, Neil, Sage
    I agree with Mark that the apache config might need some tweaking.
    Also, there are a few more configurables that will need to be updated:

    objecter_inflight_op_bytes = 0
    objecter_inflight_ops = 0

    (I think that setting these to zero will disable them, if not then
    will need to set them to higher values then the defaults)

    Specifically you can try looking at the radosgw admin socket, read the
    perf counters during a run, and see if there's any throttling going

    Another thing is that there were a bunch of performance related
    changes going into the objecter in Giant that should help concurrency.
    I did see that you mentioned that you saw performance regression
    there, which might point at some other issue with the setup.


  6. Mark Nelson
    Nov 5 (1 day ago)

    to Yehuda, Ben, me, Thiago, Peter, Kyle, John, Neil, Sage
    I wanted to say too that we've traditionally not had any 40GbE or QDR IB equipment in our lab to test RGW at these speeds. Seeing it get 3GB/s+ (at least in the better cases) is actually not terrible for a first result! We just need to figure out where we are getting held up as Yehuda mentioned.

  7. Hi All,

    Thanks everyone for your time and valuable feedback.


    There are chances of re-reading the Objects, randomization in reading is done by COSBench workload, so I just tried my best to avoid this. Yes "echo >1 " might prove good for both.

    Apache tuning was more or less based on http://httpd.apache.org/docs/2.2/mpm.html. I played around with these parameters and I believe these configs made Apache gates wide and open. I suspect FASTCGI is bigger problem here, I was also observing very high CPU usage(~95%) on RGW nodes at peak load. However I don't rule out possibility of better numbers with better tuning.

    I will try out the addition configurations recommended by you and Yahuda, will let you people know about my findings.

    I will try to attend the performance meeting.


  8. Yehuda Sadeh
    10:54 PM (13 hours ago)

    to me, Mark, Ben, Thiago, Peter, Kyle, John, Neil, Sage
    Looking more closely into the blog post I have a couple more things:

    1. You set the following apache config:

    MaxRequestsPerChild 10000

    Which basically means that each apache process dies after getting 10k
    connections. It really depends on how the benchmark is running, but it
    might have severe performance impact. You should really be setting it
    to zero.

    2. The small IO test uses relatively large IO size, beyond what the
    gateway considers as small. It may benefit with adjusting its internal
    IO params. You can add:

    rgw max chunk size = 1048576

    This will cut the number of rados IO operations in the small IO
    benchmark by half.


  9. Hi All,

    I would like to include your responses in the blog post so it will benefit open source community in general. I will share the blog post with CEPH community mailing list and have more feedback from there. FYI, I am also giving a talk about this @ OpenStack India Mini-Conf (7-Nov-2014). I will present the Ceph-RGW results with a note that there are active discussion going on .

    I hope it make sense to all.

    PS: I am working on turntables, Yehuda suggested.

  10. Hi Puspesh ,
    its Anurag, I think you gave a presentation on CEPH vs Swift Performance on OSI Days 2014 and i am interested in CEPH and trying to deploy it on simple commodity hardware just for POC purpose. but i am facing some issues and i am trying to do so inside a network with ntlm proxy using ceph-deploy . i have done it on Vm's with 2 osd and 1 monitor and its working but with physical machines i am facing problems.
    may be you can give me a email id on which we can communicate.

  11. Hey Pushpesh,

    The images for the configuration and the results do not load on this page. Can you please look into it? Otherwise, do you have a pdf file of this blog that you can email it to me?

    1. My bad I have not checked it for a while, images should work now. Study is old and I have not worked on it for long, please read it with a pinch of salt. Ceph giant release is what i worked on things might have changed for better now. Thanks for making me aware about the disappeared images.

  12. Hi,

    When a rgw service is started, by default below pools are created.

    When a swift user is created, some default pools are created. But I would like to use "Pool_A" for the swift user.
    From client when I run Cosbench the data should be placed in "Pool_A" instead of placing it in default pools. How can i achieve that.

    Also need help on how to run Cosbench with swift user. Your help is very much appreciated.


  13. This is a very nice article. thank you for publishing this. i can understand this easily.!!..Swift Online Training