Don't use big O estimates, always measure. It used to work back in the
days when double multiplication was a bottleneck. The computation cost is
effectively free on both the CPU and GPU and you're seeing pure copying
costs. Also, I'm dubious that cublas is doing what you think it is. Can you
link
For topic #4 (streaming ML in Python), there’s an existing JIRA, but progress
seems to have stalled. I’d be happy to help if you want to pick it up!
https://issues.apache.org/jira/browse/SPARK-4127
-
jeremyfreeman.net
@thefreemanlab
On Feb 26, 2015, at 4:20 PM, Xiangrui
Btw, I wish people would stop cheating when comparing CPU and GPU timings
for things like matrix multiply :-P
Please always compare apples with apples and include the time it takes to
set up the matrices, send it to the processing unit, doing the calculation
AND copying it back to where you need
I couldn't agree with you more, Sam. The GPU/Matrix guys typically don't
count their copy times, but claim that you should be doing *as much as
possible* on the GPU - so, maybe for some applications where you can
generate the data on the GPU this makes sense. But, in the context of Spark
we should
My code
```
// Initial the graph, assign a counter to each vertex that contains the
vertex id only
var anfGraph = graph.mapVertices { case (vid, _) =
val counter = new HyperLogLog(5)
counter.offer(vid)
counter
}
val nullVertex = anfGraph.triplets.filter(edge = edge.srcAttr ==
null).first
Historically, many orgs. have replaced data warehouses with Hadoop clusters
and used Hive along with Impala (on Cloudera deployments) or Drill (on MapR
deployments) for SQL. Hive is older and slower, while Impala and Drill are
newer and faster, but you typically need both for their complementary
Hi Vikram,
There was a recent presentation at Strata that you might find useful: Hive on
Spark is Blazing Fast .. Or Is
It?http://www.slideshare.net/hortonworks/hive-on-spark-is-blazing-fast-or-is-it-final
Generally those conclusions mirror my own observations: on large data sets,
Hive
FWIW, I tested the first rc and saw no regressions. I ran our benchmarks
built against spark 1.3 and saw results consistent with spark 1.2/1.2.1.
On 2/25/15, 5:51 PM, Patrick Wendell pwend...@gmail.com wrote:
Hey All,
Just a quick updated on this thread. Issues have continued to trickle
in. Not
Hey Alexander,
I don't quite understand the part where netlib-cublas is about 20x
slower than netlib-openblas. What is the overhead of using a GPU BLAS
with netlib-java?
CC'ed Sam, the author of netlib-java.
Best,
Xiangrui
On Wed, Feb 25, 2015 at 3:36 PM, Joseph Bradley jos...@databricks.com
Dean
Thanks for the info. Are you saying that we can create star/snowflake data
models using spark so they can be queried from tableau ?
On Thursday, February 26, 2015, Dean Wampler deanwamp...@gmail.com wrote:
Historically, many orgs. have replaced data warehouses with Hadoop
clusters and
There are couple things in Scala/Java but missing in Python API:
1. model import/export
2. evaluation metrics
3. distributed linear algebra
4. streaming algorithms
If you are interested, we can list/create target JIRAs and hunt them
down one by one.
Best,
Xiangrui
On Wed, Feb 25, 2015 at 7:37
Hi Steve
Thanks for the info. I will look into hivemail. Are you saying that we can
create star/snowflake data models using spark so they can be queried from
tableau ?
On Thursday, February 26, 2015, Steve Nunez snu...@hortonworks.com wrote:
Hi Vikram,
There was a recent presentation at
Okay I confirmed my suspicions of a hang. I made a request that stopped
progressing, though the already-scheduled tasks had finished. I made a
separate request that was small enough not to hang, and it kicked the hung
job enough to finish. I think what's happening is that the scheduler or the
There's no support for star or snowflake models, per se. What you get with
Hadoop is access to all your data and the processing power to build the ad
hoc queries you want, when you need them, rather than having to figure out
a schema/model in advance.
I recommend that you also ask your questions
Hi all,
I'm not surprised if the GPU is slow. It's about the bottleneck copying the
memory. Watch my talk, linked from the netlib-java github page, to
understand further. The only way to currently make use of a GPU is to do
all the operations using the GPU's kernel. You can find some prepackaged
Hi all
now, I'm trying the SparkSQL with hivecontext.
when I execute the hql like the following.
---
val ctx = new org.apache.spark.sql.hive.HiveContext(sc)
import ctx._
val queries = ctx.hql(select keyword from queries where dt =
'2015-02-01' limit 1000)
---
It seem that the number of
Hi Masaki,
I guess what you saw is the partition number of the last stage, which
must be 1 to perform the global phase of LIMIT. To tune partition number
of normal shuffles like joins, you may resort to
spark.sql.shuffle.partitions.
Cheng
On 2/26/15 5:31 PM, masaki rikitoku wrote:
Hi all
Typo - CPU was 2.5 cheaper (not GPU!)
-Original Message-
From: Ulanov, Alexander
Sent: Thursday, February 26, 2015 2:01 PM
To: Sam Halliday; Xiangrui Meng
Cc: dev@spark.apache.org; Joseph Bradley; Evan R. Sparks
Subject: RE: Using CUDA within Spark / boosting linear algebra
Evan, thank
I've had some email exchanges with the author of BIDMat: it does exactly
what you need to get the GPU benefit and writes higher level algorithms
entirely in the GPU kernels so that the memory stays there as long as
possible. The restriction with this approach is that it is only offering
high-level
Evan, thank you for the summary. I would like to add some more observations.
The GPU that I used is 2.5 times cheaper than the CPU ($250 vs $100). They both
are 3 years old. I've also did a small test with modern hardware, and the new
GPU nVidia Titan was slightly more than 1 order of magnitude
Love to hear some input on this. I did get a standalone cluster up on my
local machine and the problem didn't present itself. I'm pretty confident
that means the problem is in the LocalBackend or something near it.
On Thu, Feb 26, 2015 at 1:37 PM, Victor Tso-Guillen v...@paxata.com wrote:
Okay
The copying overhead should be quadratic on n, while the computation
cost is cubic on n. I can understand that netlib-cublas is slower than
netlib-openblas on small problems. But I'm surprised to see that it is
still 20x slower on 1x1. I did the following on a g2.2xlarge
instance with
If anyone is curious to try exporting Spark metrics to Graphite, I just
published a post about my experience doing that, building dashboards in
Grafana http://grafana.org/, and using them to monitor Spark jobs:
http://www.hammerlab.org/2015/02/27/monitoring-spark-with-graphite-and-grafana/
Code
Cool, great job☺.
Thanks
Jerry
From: Ryan Williams [mailto:ryan.blake.willi...@gmail.com]
Sent: Thursday, February 26, 2015 6:11 PM
To: user; dev@spark.apache.org
Subject: Monitoring Spark with Graphite and Grafana
If anyone is curious to try exporting Spark metrics to Graphite, I just
Of course, breakpointing on every status update and revive offers
invocation kept the problem from happening. Where could the race be?
On Thu, Feb 26, 2015 at 7:55 PM, Victor Tso-Guillen v...@paxata.com wrote:
Love to hear some input on this. I did get a standalone cluster up on my
local
25 matches
Mail list logo